Mô hình

Ling 3.1 Flash is InclusionAI's (Ant Group) hybrid-reasoning Mixture-of-Experts model (560B total, 25B active parameters, 262K-token context), a much larger successor to the 124B Ling 3.0 Flash. This OpenRouter listing is free of token cost, making it a good fit for trying out strong reasoning, coding, and agentic workloads without spend.

This is the free OpenRouter endpoint for Apodex 1.1 Mini, a reasoning-first, Apache-2.0 mixture-of-experts model (about 35B total, 3B active parameters, 262K-token context) with native function calling. It is built for long-horizon research and forecasting that works directly with files, data, code, and tools to produce verifiable results; free access is rate-limited, so it suits experimentation over production use.

Pareto 26.10 Preview is a preview build of Unbiased's proprietary multimodal composite model (text and image input, 1M-token context, up to 128K output tokens), built for research, coding, and agentic workflows. It targets frontier-level performance on general-purpose tasks at mid-tier pricing ($0.80/M input, $3.20/M output), so expect behavior to change before the next stable Pareto release.

GPT-6.1 Sol Pro is OpenAI's GPT-6.1 Sol served with reasoning mode set to "pro" for higher-quality answers on complex tasks, keeping the same 1.05M-token context and $2/$10 per-million input/output pricing. It suits difficult coding, analysis, and research problems where extra reasoning is worth more tokens and latency.

GPT-6.1 Sol is OpenAI's upgrade to GPT-6 Sol (1.05M-token context, $2/$10 per million input/output tokens), released September 29, 2026 at DevDay below the flagship GPT-6 Astra. OpenAI says it nearly matches Astra on agentic coding, computer use, and professional work at a fifth of the price, making it a strong default for coding agents and document-heavy workflows.

Claude Sonnet 5.5 is Anthropic's Sonnet-class model (1M-token context, 128K max output, $2/$10 per million input/output tokens), released September 28, 2026 as a direct upgrade to Sonnet 5 and a faster, lower-cost complement to Opus 5.5. It generates output over 30% faster with fewer tokens per task, and is strongest at well-scoped everyday work such as building features, fixing bugs, and producing polished documents, slides, and spreadsheets.

Batch-processing variant of Claude Sonnet 5.5, Anthropic's fast Sonnet-class model for everyday coding and document work. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Perceptron Mk1.5 is Perceptron's embodied reasoning model for physical agents (36K-token context, $0.15/$1.50 per million input/output tokens), released September 25, 2026. It takes text, image, video, and audio input and answers with text plus optional structured annotations (points, boxes, polygons, tracks, and clips), adding audio, video tracking, tool calling, and graded reasoning over Mk1, which makes it a fit for robots, drones, smart glasses, and other perception-driven agents.

Ember-1 is Fireworks Research's reasoning model built on Moonshot AI's Kimi K3, retrained (not a new base model) to produce reasoning traces roughly 40% shorter while keeping K3-level quality, with text and image input and a 1M-token context window. Released September 23, 2026 as a research preview, it targets coding, knowledge work, and multi-turn agent workflows where reasoning tokens drive most of the cost and latency.

GLM-5.3-Prime is the high-speed serving tier of Z.ai's open-weight GLM-5.3 flagship, running the same weights (1M-token context, up to 128K output) behind an accelerated stack for 1.5-2x the output throughput at $2.80/$8.80 per million input/output tokens. Reasoning is always on (low, high, or max effort), and it targets latency-sensitive coding, streaming code generation, and long-horizon multi-turn agent orchestration.

Qwen3.8 Max Prime is the high-speed serving tier of Alibaba's proprietary Qwen3.8 Max (same 2.4-trillion-parameter MoE weights, 1M-token context, text, image, and video input, reasoning on by default), delivering 1.5-2x the output throughput at a higher price of $4/$12 per million input/output tokens. Announced September 22, 2026 as Alibaba's "Prime mode", it is for coding, office automation, and long-running agent workflows where latency matters more than cost.

Aion 3.5 Mini is the smaller, lower-cost sibling of AionLabs' Aion 3.5, a GLM-based multi-model roleplaying and storytelling system with a 262K-token context window. It uses the same collaborative generation process as Aion 3.5 at roughly a quarter of the price, suited to high-volume or latency-sensitive roleplay and creative writing.

Aion 3.5 is AionLabs' multi-model roleplaying and storytelling system built on the GLM family (262K-token context, released September 23, 2026), in which several specialized models each contribute to every response. It succeeds Aion-3.0 and is tuned for character-driven fiction with stronger narrative structure, tension, and conflict, rather than coding or general assistant work.

Solar Mini 4 is Upstage's compact mixture-of-experts model (35B total, 3B active parameters, 524K-token context), released September 23, 2026 at $0.05/$0.20 per million input/output tokens. It is built for fast, low-cost agentic use with tool calling and structured outputs, with fluent Korean alongside strong English and Japanese, as a cheaper sibling to Solar Pro 4.

Command A+ is Cohere's flagship enterprise model, an open-weight sparse mixture-of-experts (218B total, 25B active parameters, Apache 2.0) with text and image input, 128K-token input context, and support for 48 languages, released May 20, 2026. It combines reasoning, RAG with native citations, coding, and strict tool calling in one model that runs on as few as two H100 GPUs, aimed at privately deployable agentic workflows.

GPT-6 Luna Pro is OpenAI's GPT-6 Luna served with reasoning mode set to "pro" for higher-quality answers on harder problems, keeping the same 1.05M-token context and $0.10/$0.50 per-million input/output pricing as the base model. It suits workloads that want Luna's low cost but need more careful reasoning than the default mode, at the expense of more tokens and latency per request.

This is the batch-processing variant of GPT-6 Luna Pro, OpenAI's low-cost GPT-6 Luna served with reasoning mode "pro", offered at half the standard token price. Requests run asynchronously, best suited to large-scale jobs that benefit from deeper reasoning but do not need real-time responses.

GPT-6 Luna is the fast, low-cost tier of OpenAI's GPT-6 series (1.05M-token context, text, image, and file input, $0.10/$0.50 per million input/output tokens), released September 22, 2026 below GPT-6 Sol and the flagship GPT-6 Astra. It is built for high-volume tasks with a clear goal, such as summarization, extraction, classification, quick questions, and lightweight agentic steps.

This is the batch-processing variant of GPT-6 Luna, the fast, low-cost tier of OpenAI's GPT-6 series with a 1.05M-token context window, offered at half the standard token price. Requests run asynchronously, making it a good fit for very large summarization, extraction, and classification jobs that do not need real-time responses.

GPT-6 Sol Pro is OpenAI's GPT-6 Sol served with reasoning mode set to "pro" for higher-quality answers on complex tasks, keeping the same 1.05M-token context and $2/$10 per-million input/output pricing as the base model. It targets difficult coding, analysis, and research problems where extra reasoning is worth more tokens and latency but Astra's price is not justified.

This is the batch-processing variant of GPT-6 Sol Pro, OpenAI's GPT-6 Sol served with reasoning mode "pro" for maximum quality, offered at half the standard token price. Requests run asynchronously, suited to large-scale jobs that need deeper reasoning but not real-time responses.

GPT-6 Sol is the cost-efficient high-end model in OpenAI's GPT-6 series (1.05M-token context, text, image, and file input, $2/$10 per million input/output tokens), released September 22, 2026 between the flagship GPT-6 Astra and the fast GPT-6 Luna. OpenAI reports it makes about half as many factual mistakes as GPT-5.6 Sol, reaching Astra-level reliability for everyday agentic coding and professional work at much lower cost.

This is the batch-processing variant of GPT-6 Sol, OpenAI's cost-efficient high-end GPT-6 model with a 1.05M-token context window, offered at half the standard token price. Requests run asynchronously, making it best for large coding, analysis, and document jobs that do not require real-time responses.

Claude Opus 5.5 is Anthropic's flagship model released September 22, 2026 as the first of the Claude 5.5 family, with a 1M-token context window, text, image, and file input, and pricing of $4/$20 per million input/output tokens. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5, and it leads on agentic coding, multi-step changes in large codebases, and long-horizon knowledge work.

Batch-processing variant of Claude Opus 5.5, Anthropic's Claude 5.5 flagship for agentic coding and long-horizon knowledge work with a 1M-token context window. Same capabilities as the standard model, offered at half the token price via asynchronous batch processing for large jobs that do not need real-time responses.

MiMo-V2.6-Pro-UltraSpeed is a fast-serving edition of Xiaomi's flagship MiMo-V2.6-Pro (MoE, over 1 trillion parameters, 1M-token context, omnimodal text/image/audio/video input), running the same checkpoint at roughly 10x the output speed of the standard endpoint with matching quality. It is built for latency-sensitive interactive agents that still need Pro-level coding, reasoning, and multimodal capability.

MiMo-V2.6-Flash is Xiaomi's open-source foundation model (MoE, 309B total parameters, 15B active per token, 1M-token context, omnimodal text/image/audio/video input), a lower-cost, hybrid-attention sibling to the closed flagship MiMo-V2.6-Pro. It targets high-volume workloads that need most of Pro's agentic coding and reasoning capability at a fraction of the inference cost.

MiMo-V2.6-Pro is Xiaomi's flagship foundation model (MoE, over 1 trillion parameters, 1M-token context, omnimodal text/image/audio/video input), built to push the ceiling of capability in Xiaomi's model lineup. It targets the most demanding coding, reasoning, and agentic workloads, with a faster UltraSpeed edition and a cheaper, open-weight Flash sibling available for less latency- or cost-sensitive use.

Grok 4.7 is xAI's (listed on OpenRouter under the SpaceXAI brand) proprietary flagship model, succeeding Grok 4.6, with a 500K-token context window and up to 450K output tokens. It is positioned for long-running software-engineering agents and knowledge work, with particular emphasis on verifying its own work over extended, multi-step tasks.

Qwen3.8 Omni Flash is Alibaba's omni-modal reasoning model (text, image, audio, and video input, 1M-token context, thinking on by default), released September 18, 2026 as a hosted API with no open weights. It is the first Qwen model built around agentic audio-video work such as video editing, audio-visual summarization, and real-time conversation, with speech recognition in 74 languages and very low audio input pricing.

Ternary Bonsai 2 27B is PrismML's ternary-quantized derivative of Qwen3.8-27B (27B dense parameters, Apache 2.0, 256K-token context), compressing weights to roughly 1.7 bits per parameter so the full model fits in under 9 GB while retaining most of the original's benchmark score. It targets coding, math, tool-calling, and image-understanding workloads that need to run locally on a single consumer GPU or high-end laptop rather than in the cloud.

GLM-5.3-FlashX is Z.ai's speed-optimized variant of GLM-5.3-Flash, a natively multimodal MoE model (around 320B total / 18B active parameters, MIT-licensed, 1M-token context) built on the same hybrid sparse-and-linear attention architecture. It generates at up to 200 tokens/sec for latency-sensitive coding and long-horizon agent tasks that still need the full 1M-token context, at a premium over the standard GLM-5.3-Flash endpoint.

Pareto is Unbiased's proprietary multimodal model (256K-token context, up to 128K output tokens), described by its maker as a composite system built for research, coding, and agentic workflows. It aims for frontier-level performance across general-purpose tasks, priced in the mid tier between fast, cheap models and top frontier flagships.

This is OpenRouter's auto-updating alias to DeepSeek's current top-tier 'Pro' model, always resolving to the newest release in that family (1M-token context) rather than a pinned snapshot. It lets integrations target DeepSeek's flagship reasoning line by a stable identifier without updating the model string on every new checkpoint.

This is OpenRouter's auto-updating alias to DeepSeek's current 'Flash' model, always resolving to the newest, lower-cost release in that family (1M-token context) rather than a pinned snapshot. It lets integrations track DeepSeek's newest fast, cheap model line, currently the DeepSeek V4.1 Flash generation, without updating the model string on every release.

Schematron V2 Turbo is Inference.net's 3B-parameter HTML-to-JSON extraction model (128K-token context), built on an IBM Granite 4.0 H Micro base and tuned for throughput over quality in high-volume scraping workloads. Extraction schemas are supplied via response_format rather than prompts, and it trades a little accuracy for speed against its sibling, Schematron V2 Small.

Schematron V2 Small is Inference.net's 3B-parameter HTML-to-JSON extraction model (128K-token context), built on a Llama 3.2 3B base and tuned for extraction quality on complex schemas and long pages rather than raw throughput. Like its Turbo sibling, it takes the target schema through response_format, making it a fit for accuracy-sensitive scraping and data-pipeline work.

This is OpenRouter's auto-updating alias to OpenAI's current 'Astra' model, always resolving to the newest release in that family (1.05M-token context) rather than a pinned snapshot. Astra is priced well above Sol, Terra, and Luna in the same alias family, positioning it as OpenAI's most expensive and most capable current tier on OpenRouter.

This is OpenRouter's auto-updating alias to OpenAI's current 'Sol' model, always resolving to the newest release in that family (1.05M-token context) rather than a pinned snapshot. Sol has been positioned as OpenAI's flagship-tier alias, priced below the newer, more expensive Astra alias in the same family.

This is OpenRouter's auto-updating alias to OpenAI's current 'Terra' model, always resolving to the newest release in that family (1.05M-token context) rather than a pinned snapshot. Terra sits in the mid tier by price among OpenAI's Sol/Terra/Luna/Astra alias family, between the cheaper Luna and the pricier Sol and Astra.

This is OpenRouter's auto-updating alias to OpenAI's current 'Luna' model, always resolving to the newest release in that family (1.05M-token context) rather than a pinned snapshot. Luna is the cheapest and fastest tier in OpenAI's Sol/Terra/Luna/Astra alias family, priced roughly ten times below Sol and suited to high-volume, latency-sensitive chat and lightweight agent work.

Fugu Ultra v2 is the higher-performance tier of Sakana AI's Fugu family: rather than a single trained checkpoint, OpenRouter describes it as a learned multi-agent orchestration system that routes each request across a pool of models, exposed here with a 1M-token context. It is proprietary and priced at the top of the Fugu line, aimed at the hardest coding, reasoning, and tool-use tasks where Fugu Max's lower cost is not enough.

Fugu Max is the cost-performance tier of Sakana AI's Fugu family: like Fugu Ultra v2, OpenRouter describes it as a learned multi-agent orchestration system rather than a single checkpoint, routing requests across models behind a 1M-token-context endpoint. It is proprietary and priced well below Fugu Ultra v2, fitting routine and moderately difficult coding and interactive tasks.

Ling 3.0 Flash VL is InclusionAI's open-weight multimodal MoE model (124B total / 5.5B active parameters, 128K-token context on this metered endpoint), extending the text-only Ling 3.0 Flash with native visual perception for image and video input. It is aimed at vision-grounded agentic tasks that route image and video understanding through the same reasoning pipeline used for text.

DeepSeek V4.1 Flash is DeepSeek's open-weight sparse MoE model and the first built on the company's Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output with a 1M-token context window. It targets coding, terminal and computer-use agents, and long-horizon multi-step tasks as a low-cost, high-throughput sibling to DeepSeek's Pro line.

Batch-processing variant of DeepSeek V4.1 Flash, DeepSeek's open-weight sparse MoE model built on its Causal Encoder-Decoder architecture (8B active parameters on input, 16B on output, 1M-token context). Same coding and agentic capabilities as the standard model, offered at lower cost via asynchronous batch processing for high-volume jobs that do not need real-time responses.

Mercury 2.5 is Inception Labs' diffusion-based LLM (dLLM), generating tokens in parallel rather than sequentially to reach about 1,107 tokens/second with a 260K-token context window, and is billed as the largest diffusion language model trained to date. It delivers a roughly 40% intelligence gain over Mercury 2 (comparable to cost-optimized frontier models like Claude Haiku 4.5, Gemini 3.5 Flash-Lite, and GPT-5.6 Luna Low) with tunable reasoning, parallel tool calls, and schema-aligned JSON, suiting latency-sensitive search agents, voice pipelines, and coding subagents.

Nex-N2.5-Mini is the smallest tier of Nex AGI's September 2026 Nex-N2.5 family, a multimodal agentic mixture-of-experts model (about 35B total parameters, 262K-token context) released with Apache-2.0 open weights. It targets computer use, browser automation, and agentic coding at very low per-token cost, making it a cheap default for high-volume agent loops.

Nex-N2.5-Pro is Nex AGI's multimodal agentic mixture-of-experts model (about 397B total parameters, 262K-token context), the middle tier of the September 2026 Nex-N2.5 family above Mini. It builds on Nex-N2 with focused gains in computer use, web browsing, and visually grounded agentic coding, suiting long-horizon agent workloads that need a visual feedback loop.

GPT-6 Astra is OpenAI's flagship GPT-6-generation model (1.05M-token context, up to 128K output tokens, priced at $10/$50 per million input/output tokens), built for demanding end-to-end work. It excels at advanced analysis, software engineering, deep research, scientific work, and document creation, with particular strength in long-horizon agentic tasks involving computer and browser use.

This is the batch-processing variant of OpenAI's flagship GPT-6 Astra, offering the same 1.05M-token context and long-horizon agentic, coding, and research capabilities at lower cost for asynchronous workloads. Requests run on a delayed, non-interactive basis, making it best for large-scale jobs that do not require real-time responses.

GPT-6 Astra Pro is OpenAI's flagship GPT-6 Astra served with reasoning mode set to "pro" for higher-quality answers on the hardest tasks, keeping the same 1.05M-token context window and $10/$50 per-million input/output pricing as the base model. It targets demanding end-to-end work such as advanced analysis, software engineering, deep research, scientific work, and long-horizon agentic tasks involving computer and browser use, trading extra latency and reasoning cost for maximum quality.

This is the batch-processing variant of GPT-6 Astra Pro, OpenAI's flagship GPT-6 Astra served with reasoning mode "pro" for maximum quality, offered at lower cost for asynchronous workloads. Requests run on a delayed, non-interactive basis, best suited to large-scale jobs that do not need real-time responses.

This is the free OpenRouter endpoint for InclusionAI's (Ant Group) Ling-3.0-Flash-Sante, a health and medicine fine-tune of the Ling 3.0 Flash base (124B total, about 5.1B active parameters, 262K-token context), released September 4, 2026. It targets medical knowledge reasoning, clinical safety, and evidence-based retrieval (leading its size class on benchmarks like MedXpertQA-Text and DiagnosisArena-MCQ) while retaining general reasoning and coding ability; free access is rate-limited and best suited for evaluation rather than production.

Qwen3.8-Max-0902 is Alibaba's September 2, 2026 snapshot of its proprietary flagship Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with a 1M-token context window that accepts text, image, and video input and has reasoning on by default. The architecture and pricing are unchanged from the base model; this post-training update sharpens coding and agentic performance across multi-step software projects, multi-tool orchestration, and long-horizon task execution.

This is the discounted 'contributor' tier of Meta's Muse Spark 1.3 coding and agentic model, priced far below the standard endpoint (about $0.10/$0.20 per million input/output tokens versus $1.25/$4.25) in exchange for letting Meta use your prompts and completions for training. Capabilities match the standard model, so it is best reserved for non-sensitive workloads where that data-sharing tradeoff is acceptable.

Muse Spark 1.3 is Meta's proprietary coding and agentic model, released September 2, 2026 with a 1M-token context window and text, image, and video input, served through Muse Code and the Meta Model API. It is Meta's largest coding and agentic improvement to date, running roughly 20% fewer tool calls and 25% fewer tokens per task than Muse Spark 1.2 while sustaining longer-horizon work, and defaults to its broadly available xhigh reasoning level.

Gemini 3.8 Flash is Google's workhorse Flash model, released September 2, 2026 and built on Gemini 3.7 Flash, with a 1M-token context window, 64K max output, tunable thinking levels, and multimodal input across text, image, audio, video, and PDF. It keeps the same context window and pricing as its predecessor while notably improving coding and reasoning and reducing hallucinations, making it a strong low-latency choice for high-volume agentic and everyday tasks.

Gemini 3.8 Flash is Google's workhorse Flash model built on Gemini 3.7 Flash, with a 1M-token context window, tunable thinking levels, and multimodal input, tuned for low-latency agentic and everyday tasks. This is the batch variant for asynchronous, lower-cost processing.

Claude Fable 5.1 is Anthropic's most capable model, released September 2026 as an upgrade to Claude Fable 5 for coding, knowledge work, and long-running problem-solving, with a 1M-token context window. It improves agentic coding by over 30% and doubles its predecessor's Terminal-Bench-Science score while cutting costs through cheaper cache reads, and is the openly available counterpart to the more tightly safeguarded Mythos 5.1.

Batch-processing variant of Claude Fable 5.1, Anthropic's most capable model for coding, knowledge work, and long-running problem-solving with a 1M-token context window. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Granite 4.2 8B is IBM's dense, decoder-only 8B-parameter reasoning model (Apache 2.0 license, 128K native context extendable to 512K) with switchable full, low-effort, and non-thinking modes, released August 25, 2026 as the mid-size member of the Granite 4.2 family. IBM positions it for enterprise use in mathematics, code generation, multilingual dialogue across 12 languages, and agentic tool-calling workflows on modest hardware.

Hy4 Preview is Tencent Hunyuan's open-weight mixture-of-experts flagship (770B total parameters, 49B active, Apache 2.0 license, 1M-token context), released August 28, 2026 as a preview ahead of a later official release. It targets coding agents, complex tool-use workflows, and productivity tasks like office work, game development, and scientific research rather than general chat.

Ling-3.0-flash-fin is inclusionAI's (Ant Group) finance-tuned variant of Ling-3.0-flash, an open-weight 124B-parameter mixture-of-experts model (about 5.1B active per token, 256K-token context). It is optimized for financial research and multi-step investment workflows with heavy tool use, while retaining the base model's general reasoning, coding, and math abilities.

This is an alias that always points to Z.ai's latest GLM Flash model.

Qwen3.8 Flash is Alibaba's speed- and cost-optimized model in the Qwen3.8 family (1M-token context, up to 131K output tokens) with native text, image, and video input plus tool calling and structured outputs, released August 26, 2026. It sits below the flagship Qwen3.8 Max and is aimed at coding assistance, agentic workflows, document and codebase analysis, and long-video understanding where speed and cost matter more than peak capability.

GLM-5.3-Flash is Z.ai's first natively multimodal model in the GLM-5 family (320B total parameters, 18B active mixture-of-experts, MIT license, 1M-token context), released August 26, 2026 with a hybrid sparse-plus-linear attention architecture that keeps long-context behavior accurate while cutting compute overhead. It is built for efficient coding and long-horizon agent tasks, outperforming GLM-5.2 at roughly one-tenth the price while approaching flagship-class coding and agentic benchmark performance.

This is the batch-processing variant of GLM-5.3-Flash, Z.ai's open-weight natively multimodal mixture-of-experts model (320B total parameters, 18B active, MIT license, 1M-token context, hybrid sparse-plus-linear attention), offering the same coding and agentic capabilities at a lower cost via OpenRouter's asynchronous batch endpoint. It is best for high-volume, non-latency-sensitive coding or agent workloads where cost efficiency matters more than immediate response time.

Muse Spark 1.2 Contributor is Meta Superintelligence Labs' cost-reduced access tier for its proprietary Muse Spark 1.2 reasoning model (a Stable LatentMoE design activating 16 of 896 experts per token via Kimi Delta Attention), offering a roughly 1M-token context window across text, image, video, audio, and PDF input. In exchange for pricing more than 10 times cheaper than the standard tier, prompts and outputs on this tier may be used by Meta to improve its products, making it best suited for experimentation and early-stage projects rather than data-sensitive production use.

DeepSeek V4 Flash Vision Exp is an experimental multimodal variant of DeepSeek V4 Flash 0731, a sparse MoE model (284B total, 13B active parameters) with a 1M-token context window that now accepts image input alongside text. It matches V4 Flash on agentic reasoning and world knowledge while adding document and chart understanding, visual question answering, and multimodal agent workflows, positioning it as the multimodal entry in DeepSeek's lightweight Flash line.

Hy-MT2-1.8B is Tencent Hunyuan's smallest open-weight translation model (1.8B dense parameters, Apache 2.0 license), supporting 33 languages as the lightest of the three Hy-MT2 'fast-thinking' sizes alongside the 7B dense and 30B-A3B MoE variants. Despite its size it beats mainstream commercial translation APIs on general and business-domain benchmarks, and AngelSlim quantization down to about 440MB makes it deployable on mobile chipsets for on-device translation.

Hy-MT2-30B-A3B is Tencent Hunyuan's largest translation model, a sparse mixture-of-experts architecture (30B total parameters, 3B active, 128 routed plus 1 shared expert) supporting 33 languages, released open-weight (Apache 2.0) alongside the smaller 1.8B and 7B dense Hy-MT2 variants. In fast-thinking mode it outperforms larger open-source general models such as DeepSeek-V4-Pro and Kimi K2.6 on translation benchmarks, making it the pick for high-quality business and domain-specific translation where the dense variants fall short.

This is an alias that always points to Z.ai's latest GLM model.

Hy-MT2-7B is Tencent's 7B-parameter translation model (8,192-token context), released August 19, 2026 as the mid-size entry in the Hy-MT2 fast-thinking translation family alongside 1.8B and 30B-A3B MoE siblings. It supports 33 language pairs plus Chinese dialect and minority-language pairs, with structured, delimiter-based, contextual, glossary-based, and style-guided translation workflows for dedicated multilingual translation rather than general-purpose chat.

GLM-5.3 is Z.ai's open-weight flagship model (743B parameters, 1M-token context), an incremental upgrade over GLM-5.2 with stronger coding and much better token efficiency per task. Reasoning is always on with low, high, and max effort levels (max by default), and Z.ai positions it as a top open-weight model for complex software engineering and long-horizon agent tasks.

This is the batch-processing variant of Z.ai's open-weight GLM-5.3 flagship model, offering the same 743B-parameter, 1M-token coding and long-horizon agentic capabilities at lower cost for asynchronous workloads.

Qwen3.8 27B is a smaller, Apache 2.0-licensed open-weight model from Alibaba's Qwen3.8 generation, suited for general-purpose chat, coding, and agentic tasks at moderate cost.

Dots3-Note Preview is an open-weight mixture-of-experts model from Dots Studio (280B total, 16B active) with a 512K token context window and multimodal input support. It is the lightest model in the dots3 family, suited for coding, long-context document processing, multimodal understanding, and multi-step agentic workflows; this is the free-tier variant.

Gemini 3.7 Flash is Google's most intelligent workhorse model yet for coding and agents, an iterative refinement of 3.6 Flash with notably higher accuracy on debugging and issue resolution at roughly half the cost. It is designed for responsive multi-step agentic workflows, coding, and everyday knowledge work.

Gemini 3.7 Flash is Google's most intelligent workhorse model yet for coding and agents, an iterative refinement of 3.6 Flash with notably higher accuracy on debugging and issue resolution. This is the batch variant for asynchronous, lower-cost processing.

Seed 2.1 Turbo is a multimodal model from ByteDance Seed built for coding and long-horizon agent workflows, including end-to-end software delivery, multi-step task execution, and video/image understanding. It supports planning, debugging, and self-correction with a 262K-token context window.

Qwen3.8 2.4T-A95B is Alibaba's massive open-weight mixture-of-experts flagship (2.4 trillion total parameters, 95B active), representing the top of the Qwen3.8 generation for frontier-level reasoning, coding, and agentic tasks.

Seed 2.0 Code is ByteDance Seed's model optimized specifically for agentic coding, suited to frontend development and multilingual programming tasks. It is designed to work well inside coding-agent tools such as Claude Code, Kilo, and OpenCode.

DeepSeek V4 Pro (0813) is the August 13, 2026 general-availability checkpoint of DeepSeek V4 Pro, adding DSpark speculative decoding, reasoning-effort levels, and native OpenAI Responses API support. It is DeepSeek's flagship model for coding, reasoning, and agentic tasks in the V4 generation.

Grok 4.6 is xAI's flagship model built for long-running agentic work and ambitious interactive and visual tasks, with a 500,000-token context window and text/image input. It is tuned for advanced reasoning, coding, and complex agent workflows at the top of xAI's model lineup.

LFM2.5-2.6B is Liquid AI's on-device agentic model (2.6B parameters, 128K context), pre-trained on ~34T tokens to plan, call tools, and run multi-step tasks entirely on phones, laptops, and edge devices with no data leaving the device. It leads its size class on instruction-following and tool-use benchmarks while running in under 2.5 GB of memory; this is OpenRouter's rate-limited free endpoint.

Nemotron 3.5 Lightning is NVIDIA's open 30B mixture-of-experts model with 3B active parameters and a 1M-token context window, built for fast, accurate execution within long-running agent pipelines. It is designed as a low-latency, low-cost execution layer for always-on agentic systems.

Nemotron 3.5 Lightning is NVIDIA's open 30B mixture-of-experts model with 3B active parameters and a 1M-token context window, built as a low-latency, low-cost execution layer for always-on agentic systems. This is the free-tier variant of the same model.

Sakana Namazu is an OpenAI-compatible model from Sakana AI, built as a fine-tune of Moonshot AI's open-weight Kimi model and tailored for Japanese language and business use cases. It suits Japanese-language chat, business writing, and localized assistant tasks.

Solar Pro 4 is Upstage's next-generation flagship model, an agent-first successor to Solar Pro 3 focused on further reasoning gains at low cost. It targets multi-step tool use and production-scale agentic workloads while retaining Upstage's efficiency-focused, compact-parameter design philosophy.

Muse Glimmer 30B is Meta's open-weight multimodal model, a distilled version of the larger Muse Spark model, combining an LLM with a dedicated perception encoder for text and visual input. It is built for local agentic and coding workloads, and can run entirely offline on a single consumer GPU.

Muse Spark 1.2 is an updated release in Meta Superintelligence Labs' Muse family, sharing Muse Spark's focus on multimodal reasoning, coding, and AI-assisted software development with long-context support. It offers incremental improvements over Muse Spark 1.1 for agentic and coding workflows.

This is an alias that always points to DeepSeek's latest V4-Flash model.

DeepSeek V4 Flash (0731) is the July 31, 2026 public-beta checkpoint of DeepSeek's V4 Flash model, DeepSeek's fast, cost-efficient V4-generation model tuned for agentic workflows. It offers the same lightweight, high-throughput design as other V4 Flash releases.

Inkling-Small is a compact open-weights sibling of Thinking Machines Lab's Inkling model, at 276B total parameters (12B active) versus Inkling's 975B. It reasons natively over text, images, and audio with adjustable thinking effort, achieving performance close to the full Inkling model at a fraction of its size.

This is the free OpenRouter endpoint for Thinking Machines Lab's Inkling-Small, a 276B-parameter mixture-of-experts model (12B active, Apache 2.0 license, 1M-token context) released July 30, 2026 as the smaller, efficient sibling in the Inkling family. Free access is rate-limited to agentic-harness use with prompt and output logging, making it best suited for evaluating its coding, agentic, and RAG capabilities rather than production workloads.

Qwen3.7 Flash is Alibaba's fast, low-cost proprietary tier of the API-only Qwen3.7 generation, built for high-throughput, latency-sensitive tasks.

Claude Opus 5 is Anthropic's flagship model in the Claude 5 generation, representing the top of Anthropic's intelligence tier for coding, agentic workflows, and complex reasoning. It targets the most demanding professional software engineering and long-horizon agent tasks Anthropic offers.

Batch-processing variant of Claude Opus 5, Anthropic's Claude 5 generation flagship model. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Ling 3.0 Flash is InclusionAI's (Ant Group) successor to Ling 2.6 Flash, a hybrid-reasoning Mixture-of-Experts model (~124B total, ~5.1B active) combining Kimi Delta Attention with Multi-Head Latent Attention for efficient long-range memory. It supports both thinking and non-thinking modes with a native 262K context, suited for fast general-purpose coding and agentic tasks.

Laguna S 2.1 is an open-weight agentic coding model from Poolside, a 118B-parameter mixture-of-experts system with 8B active parameters per token and up to a 1M-token context window. Trained with reinforcement learning inside Poolside's own agent harness, it targets agentic software engineering tasks and is competitive with much larger models on SWE-bench and Terminal-Bench.

Free-tier variant of Poolside's Laguna S 2.1, an open-weight 118B MoE agentic coding model with a 1M-token context window trained via RL in Poolside's agent harness. Same underlying model as the paid version, suited for agentic software engineering and terminal/coding tasks.

Gemini 3.6 Flash is Google's workhorse model delivering a step up in coding, knowledge work, and multimodal performance over 3.5 Flash, with meaningfully improved token efficiency. It targets everyday agentic and coding tasks that need strong quality at Flash-tier cost.

Gemini 3.6 Flash is Google's workhorse model delivering a step up in coding, knowledge work, and multimodal performance with improved token efficiency. This is the batch variant for asynchronous, lower-cost processing.

Gemini 3.5 Flash-Lite is the most cost-effective model in Google's 3.5 Flash line, built for high-volume, low-latency tasks where price and speed are the priority. It trades peak reasoning depth for efficiency at scale.

Gemini 3.5 Flash-Lite is the most cost-effective model in Google's 3.5 Flash line, built for high-volume, low-latency tasks where price and speed are the priority. This is the batch variant for asynchronous, lower-cost processing.

LongCat-2.0 is Meituan's open-weight 1.6-trillion-parameter Mixture-of-Experts model (~48B active) with a native 1M-token context window, using sparse attention for efficient long-context processing. It targets agentic coding, repository-level code changes, and long-horizon problem solving, with reported performance comparable to Gemini 3.1 Pro on coding benchmarks.

Inkling is Thinking Machines Lab's first flagship model, a 975B-parameter (41B active) open-weights mixture-of-experts model that reasons natively over text, images, and audio with up to a 1M-token context window. It is designed as a general-purpose frontier model for multimodal reasoning and agentic tasks, released under Apache 2.0.

This is the free OpenRouter endpoint for Thinking Machines Lab's Inkling, a 975B-parameter mixture-of-experts model (41B active, Apache 2.0 license, 1M-token context) released July 2026 as the flagship of the Inkling family and the company's first in-house model. Free access is rate-limited to agentic-harness use with prompt and output logging, making it best suited for evaluating general-purpose reasoning, coding, agentic tool-use, and RAG rather than production workloads.

Auto Router (Beta) is the early-access track of OpenRouter's task-aware Auto Router, where new routing behaviors land before they reach the stable openrouter/auto. It uses the same market-driven selection (prompts classified into ~30 task types, routed by trailing 7-day community spend with a cost_tier dial) and prices each response at the routed model's rate; choose it to try the newest routing changes first.

Kimi K3 is Moonshot AI's July 2026 flagship model, a 2.8-trillion-parameter mixture-of-experts model with native multimodal vision and a 1M-token context window. It is built for long-horizon coding, reasoning, and agent workflows, positioned as an open-weight competitor to top proprietary frontier models.

This is the asynchronous batch-processing endpoint for Moonshot AI's Kimi K3, a 2.8T-parameter open-weight mixture-of-experts reasoning model (1M-token context) released July 2026 under a custom revenue-tiered Kimi K3 License. Unlike typical batch endpoints this one carries no price discount over the standard endpoint, so it is worth reaching for only when queued, asynchronous processing itself is the requirement rather than for cost savings.

Muse Spark 1.1 is a large language model from Meta Superintelligence Labs, the first model in Meta's Muse family, designed for multimodal reasoning, coding, and AI-assisted software development with up to a million tokens of context. It targets complex agentic and coding workflows requiring long-context understanding.

KAT-Coder-Pro V2.5 is Kwaipilot's flagship agentic coding model, trained on 100,000+ verifiable repository environments and capable of autonomously handling an entire issue or business workflow end to end in a real codebase. It ranks highly on agentic coding benchmarks like PinchBench and SWE-Bench Pro.

GPT-5.6 Luna Pro is a higher-reasoning-effort version of the Luna tier in OpenAI's GPT-5.6 family, aimed at latency-sensitive chat, classification, and lightweight agentic tasks that need a bit more accuracy than the base Luna model.

GPT-5.6 Luna Pro is a higher-reasoning-effort version of the fast, low-cost Luna tier in OpenAI's GPT-5.6 family. This is the batch variant for cheaper asynchronous processing.

GPT-5.6 Luna is the fast, low-cost tier of OpenAI's GPT-5.6 family, tuned for high-volume, latency-sensitive work like chat, classification, and lightweight agents while retaining genuine GPT-5.6 reasoning.

GPT-5.6 Luna is the fast, low-cost tier of OpenAI's GPT-5.6 family for high-volume chat, classification, and lightweight agents. This is the batch variant for cheaper asynchronous processing.

GPT-5.6 Terra Pro is a higher-reasoning-effort version of the balanced Terra tier in OpenAI's GPT-5.6 family, for everyday coding, reasoning, and agentic tasks that need a bit more accuracy.

GPT-5.6 Terra Pro is a higher-reasoning-effort version of the balanced Terra tier in OpenAI's GPT-5.6 family. This is the batch variant for cheaper asynchronous processing.

GPT-5.6 Terra is the balanced, everyday tier of OpenAI's GPT-5.6 family, offering coding, reasoning, and agentic performance competitive with GPT-5.5 at roughly half the cost.

GPT-5.6 Terra is the balanced, everyday tier of OpenAI's GPT-5.6 family, offering coding, reasoning, and agentic performance at low cost. This is the batch variant for cheaper asynchronous processing.

GPT-5.6 Sol Pro is an extended-reasoning version of the Sol flagship tier in OpenAI's GPT-5.6 family, aimed at the hardest complex reasoning, multi-step coding, and long-horizon agentic tasks.

GPT-5.6 Sol Pro is an extended-reasoning version of the Sol flagship tier in OpenAI's GPT-5.6 family, for the hardest reasoning and agentic coding tasks. This is the batch variant for cheaper asynchronous processing.

GPT-5.6 Sol is the flagship tier of OpenAI's GPT-5.6 family, built for complex reasoning, multi-step coding, and long-horizon agentic work, with particular strength on command-line and terminal-driven tasks.

GPT-5.6 Sol is the flagship tier of OpenAI's GPT-5.6 family, built for complex reasoning, multi-step coding, and long-horizon agentic work. This is the batch variant for cheaper asynchronous processing.

Grok 4.5 is a flagship xAI model positioned for code generation and general high-intelligence use cases, sitting between Grok 4.3 and Grok 4.6 in xAI's rapid release cadence. It is designed for demanding reasoning, coding, and agentic workloads.

This is an alias that always points to xAI's latest flagship Grok model.

Aion-3.0-mini is a smaller, faster member of AionLabs' Aion-3.0 roleplay and storytelling family, built on DeepSeek models with a similar collaborative multi-model generation approach. It targets the same creative-writing and roleplay use cases at lower cost and latency.

Aion-3.0 is a multi-model roleplaying and storytelling system from AionLabs built on the GLM family, using a collaborative generation process where multiple specialized models contribute to each response. It is designed for rich narrative structure and compelling character-driven fiction.

Hy3 is Tencent Hunyuan's open-source flagship model, a 295B-parameter MoE (about 21B active) with a hybrid fast/slow reasoning system and 256k context. It targets coding, research, and agentic product integrations, emphasizing reliable tool calling and long multi-turn conversations.

Laguna XS 2.1 is Poolside's smaller open-weight coding model (33B total parameters MoE), built for local, lightweight agentic coding workloads. It applies learnings from Poolside's larger Laguna models while staying fast and cheap enough to run on modest hardware.

Free-tier variant of Poolside's Laguna XS 2.1, a compact open-weight MoE coding model designed for local agentic coding. Same underlying model as the paid version, aimed at lightweight, cost-effective code generation and tool-use tasks.

Claude Sonnet 5 is Anthropic's most capable Sonnet-tier model, bringing near-Opus intelligence to coding, agentic workflows, and professional knowledge work at Sonnet-level cost. It is a strict improvement over Sonnet 4.6 across the cost-performance curve and serves as the default model for Anthropic's Free and Pro plans.

Batch-processing variant of Claude Sonnet 5, Anthropic's near-Opus-capable Sonnet-tier flagship. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Sakana Fugu Ultra is Sakana AI's multi-agent orchestration model that coordinates a pool of frontier language models through a single API endpoint, dynamically assigning roles via learned orchestration. It targets complex, multi-step tasks like agentic software engineering, achieving top-tier benchmark scores by intelligently routing work across underlying models.

North Mini Code is Cohere's first developer-focused coding model, a 30B-total/3B-active mixture-of-experts model released open-weight under Apache 2.0. It is trained for agentic coding, terminal tasks, and tool use across agent harnesses like OpenCode and SWE-Agent, with a 256K context window; this is the free-tier variant of the same model.

GLM-5.2 is Z.ai's flagship open-source model, shipping with a 1M-token context window, two selectable reasoning effort levels, and an unrestricted MIT license. It is competitive with top closed models on demanding software-engineering benchmarks, trailing Claude Opus-class models only narrowly while outperforming several rivals on coding tasks.

Kimi K2.7 Code is Moonshot AI's June 2026 specialized coding model, tuned for stronger instruction-following over long contexts and higher success on coding tasks. It cuts thinking-token usage by about 30% versus K2.6 while improving benchmark scores, making it well suited for agentic coding workflows.

This is an alias that always points to Anthropic's latest Claude Fable model.

Claude Fable 5 is Anthropic's most capable coding-focused model, built for long-running, ambiguous, asynchronous work such as legacy migrations and gnarly production bugs that can span hours or days. It self-corrects through verification loops and is state of the art on nearly all tested coding benchmarks, with a 1M-token context window.

Batch-processing variant of Claude Fable 5, Anthropic's most capable coding model for long-running, ambiguous, asynchronous engineering work. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Nemotron 3.5 Content Safety is a small NVIDIA moderation model built on Google's Gemma-3-4B and fine-tuned on multimodal, multilingual safety data. It classifies prompts and responses across 23 safety categories such as violence, hate, self-harm, and PII in 12 languages, and is meant for guardrailing other models rather than general chat.

Nemotron 3.5 Content Safety is a small NVIDIA moderation model built on Google's Gemma-3-4B and fine-tuned on multimodal, multilingual safety data. It classifies prompts and responses across 23 safety categories such as violence, hate, self-harm, and PII in 12 languages, and is meant for guardrailing other models rather than general chat.

Nemotron 3 Ultra is NVIDIA's flagship open-weight model in the Nemotron 3 family, a 550B-parameter mixture-of-experts model with 55B active parameters. It is optimized for orchestrating complex, long-running agent workflows, combining frontier reasoning with high inference throughput.

Nemotron 3 Ultra is NVIDIA's flagship open-weight model in the Nemotron 3 family, a 550B-parameter mixture-of-experts model with 55B active parameters optimized for complex, long-running agent workflows. This is the free-tier variant of the same model.

Qwen3.7 Plus is Alibaba's multimodal agent-focused model in the API-only Qwen3.7 generation, combining vision and language understanding with strong agentic tool-use capability.

MiniMax-M3 is MiniMax's latest M-series model for agentic reasoning, tool use, coding, multimodal chat input, and long-context tasks. It represents the newest generation of MiniMax's flagship line, building on the coding and agentic strengths of the M2 series.

Step 3.7 Flash is StepFun's multimodal vision-language model, built on Step 3.5 Flash with an added vision encoder for native image understanding. With a 256k context window, selectable reasoning levels, and high throughput, it targets agentic, coding, and search workflows that mix text and visual input.

Claude Opus 4.8 is Anthropic's most capable generally available model in the Opus family at this point in the lineup, supporting text, image, and file input with reasoning and a 1M-token context window. It is aimed at the highest-difficulty coding, analysis, and agentic tasks.

Batch-processing variant of Claude Opus 4.8, Anthropic's most capable generally available Opus model. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Qwen3.7 Max is Alibaba's flagship API-only model (no open weights) in the Qwen3.7 generation, positioned as the text flagship for advanced reasoning and agentic tasks.

Grok Build 0.1 is xAI's fast coding-specialist model, trained specifically for agentic software-engineering workflows like code generation, review, debugging, and multi-step development rather than general chat. It has a 256K context window with text and image input and is optimized for interactive coding agents and high-volume tool-use workloads.

Gemini 3.5 Flash is Google's near-Pro intelligence model at Flash-tier cost and speed, delivering strong coding proficiency and parallel agentic execution for complex, long-horizon tasks. It is designed to give developers frontier-level agent and coding performance without Pro-level pricing.

Gemini 3.5 Flash is Google's near-Pro intelligence model at Flash-tier cost and speed, delivering strong coding proficiency and parallel agentic execution for complex, long-horizon tasks. This is the batch variant for asynchronous, lower-cost processing.

Perceptron Mk1 is the flagship vision-language model from Perceptron AI, a startup founded by former Meta FAIR researchers, built for video understanding and embodied/spatial reasoning. It is designed for video QA, summarization and event detection, OCR and document parsing on messy real-world inputs, object detection/counting, and pose estimation, delivering frontier-competitive results at a much lower cost.

Gemini 3.1 Flash-Lite is Google's high-efficiency multimodal model optimized for low-latency, high-volume workloads across text, image, video, audio, and PDF input. It is designed for lightweight agentic workflows and simple data extraction where responsiveness and cost matter most.

Gemini 3.1 Flash-Lite is Google's high-efficiency multimodal model optimized for low-latency, high-volume workloads across text, image, video, audio, and PDF input. This is the batch variant for asynchronous, lower-cost processing.

GPT Chat Latest is an alias that always points to the current production ChatGPT model snapshot, tuned for natural conversational chat rather than heavy extended reasoning. It suits general assistant use where staying in sync with ChatGPT's live behavior matters.

Grok 4.3 is an iteration in xAI's Grok 4.x line, improving on Grok 4.20 with incremental gains in reasoning, coding, and agentic task performance. It is positioned as a high-intelligence general-purpose model for complex problem solving.

This is the batch-processing variant of xAI's Grok 4.3, offering the same high-intelligence reasoning, coding, and agentic capabilities of the Grok 4.x line at lower cost for asynchronous workloads. Requests are processed on a delayed, non-interactive basis, making it best for large-scale jobs that do not need real-time responses.

Mistral Medium 3.5 is Mistral AI's April 2026 flagship-tier model, a dense 128B model with a 256K context window that merges instruction-following, reasoning, and coding into a single set of weights. It is the recommended successor to Mistral Medium 3.1 for general-purpose and agentic tasks.

This is the batch-processing variant of Mistral Medium 3.5 (mistral-medium-3-5), Mistral AI's 128B-parameter dense model (Modified MIT license, 256K-token context) that merges the former Magistral reasoning and Devstral 2 coding models into one self-hostable system. It targets balanced, flagship-adjacent agentic and coding workloads with reliable multi-tool calling, and the batch endpoint trades real-time latency for lower cost on asynchronous, high-volume jobs.

Nemotron 3 Nano Omni Reasoning is a multimodal, reasoning-tuned variant of NVIDIA's Nemotron 3 Nano, built for omni (text and image, likely audio) understanding combined with extended chain-of-thought. It targets efficient agentic applications that need both reasoning and multimodal input at low compute cost.

This is an alias that always points to Anthropic's latest Claude Haiku model.

This is an alias that always points to OpenAI's latest smaller/cheaper GPT model.

This is an alias that always points to Google's latest Gemini Pro model.

This is an alias that always points to Moonshot AI's latest Kimi model.

This is an alias that always points to Google's latest Gemini Flash model.

This is an alias that always points to Anthropic's latest Claude Sonnet model.

An April 2026 snapshot of Qwen3.5 Plus, Alibaba's proprietary mid-tier model in the Qwen3.5 generation, offering the same general-purpose reasoning and agentic capability as other Plus-tier Qwen3.5 releases.

Qwen3.6 Flash is Alibaba's fast, low-cost proprietary tier of the Qwen3.6 generation, aimed at high-throughput, latency-sensitive chat and agentic tasks.

Qwen3.6 35B-A3B is a sparse mixture-of-experts model from Alibaba's Qwen3.6 line (35B total, 3B active parameters), offering efficient general-purpose and agentic capability at low inference cost.

Qwen3.6 Max Preview is a preview of Alibaba's flagship proprietary Qwen3.6 model, intended for top-tier reasoning, coding, and agentic performance ahead of general availability.

Qwen3.6 27B is an open-weight model from Alibaba's Qwen3.6 generation, a successor to Qwen3.5 with incremental improvements to general chat, reasoning, and agentic performance.

GPT-5.5 Pro is the extended-reasoning tier of GPT-5.5, for the hardest coding, research, and analytical tasks where maximum accuracy is worth the extra latency and cost.

GPT-5.5 Pro is the extended-reasoning tier of GPT-5.5 for the hardest coding, research, and analytical tasks. This is the batch variant for cheaper asynchronous processing.

GPT-5.5 is OpenAI's April 2026 flagship model, described as its most intuitive yet at understanding intent and carrying more of a task itself. It excels at writing and debugging code, research, data analysis, and multi-step agentic tool use, with reduced hallucination in sensitive domains like law, medicine, and finance.

GPT-5.5 is OpenAI's April 2026 flagship model, excelling at coding, research, data analysis, and agentic tool use with reduced hallucination in sensitive domains. This is the batch variant for cheaper asynchronous processing.

DeepSeek V4 Pro is the flagship variant of DeepSeek's V4 model family, featuring reasoning-effort levels (low/high/max) and strong performance across coding, reasoning, and agentic benchmarks. It is DeepSeek's top-tier general-purpose and reasoning model in the V4 generation.

DeepSeek V4 Flash is the faster, cheaper variant of DeepSeek's V4 model family, released in public beta and reported to substantially exceed V4-Pro-Preview on agent benchmarks. It is designed for high-throughput agentic and coding workloads where speed and cost matter alongside strong capability.

Hy3-preview is the early-access preview release of Tencent Hunyuan's Hy3 model, a large MoE architecture with hybrid fast/slow reasoning aimed at coding, research, and agentic workflows. It offers the same general capabilities as the final Hy3 release under a more restrictive preview license.

MiMo v2.5 Pro is the larger, higher-capability variant of Xiaomi's MiMo v2.5 open-source reasoning model line. It targets more demanding reasoning, math, and coding tasks than the base MiMo v2.5 model while retaining an open-source, efficiency-conscious design.

MiMo v2.5 is part of Xiaomi's MiMo family of open-source reasoning models, following the original MiMo-7B focused on math, code, and reasoning-heavy tasks. It scales up Xiaomi's reasoning-focused training approach for stronger multimodal and general reasoning performance.

This is an alias that always points to Anthropic's latest Claude Opus model.

Kimi K2.6 is Moonshot AI's April 2026 open-weight model that ties GPT-5.5 on coding benchmarks like SWE-Bench Pro while costing roughly 80% less per token. It is designed for high-end coding, reasoning, and agentic tasks at a fraction of proprietary-model cost.

Claude Opus 4.7 is the next generation of Anthropic's Opus family, purpose-built for long-running, asynchronous agents that operate with minimal supervision. It carries a 1M-token context window and targets complex, multi-step engineering and research tasks.

Batch-processing variant of Claude Opus 4.7, Anthropic's Opus model built for long-running asynchronous agents. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

GLM-5.1 is an incremental upgrade to Z.ai's GLM-5 flagship model, released as open source in April 2026 with improved reasoning and agentic performance. It continues GLM-5's focus on complex systems design and large-scale programming tasks.

Gemma 4 26B-A4B IT is a Mixture-of-Experts model from Google's open-weight Gemma 4 family with 26B total but only ~4B active parameters, giving high-throughput inference for advanced reasoning tasks. It is a multimodal, hybrid-thinking model supporting 140+ languages and up to 256K context.

Gemma 4 26B-A4B IT is a Mixture-of-Experts model from Google's open-weight Gemma 4 family with 26B total but only ~4B active parameters, giving high-throughput inference for advanced reasoning tasks. This is the free-tier variant of the same model.

Gemma 4 31B IT is a dense model from Google's open-weight Gemma 4 family that bridges server-grade performance and local execution, ranking near the top of open-model leaderboards. It handles multimodal, hybrid-thinking reasoning across 140+ languages with up to 256K context, well suited for advanced chat, coding, and reasoning.

Gemma 4 31B IT is a dense model from Google's open-weight Gemma 4 family that bridges server-grade performance and local execution, ranking near the top of open-model leaderboards. This is the free-tier variant of the same model.

Qwen3.6 Plus is Alibaba's proprietary mid-tier model in the Qwen3.6 generation, balancing cost and capability for general chat, reasoning, and agentic use cases.

GLM-5V-Turbo is a fast, lower-latency vision-language variant in Z.ai's GLM-5 family, combining multimodal image/video understanding with efficient serving. It is designed for agentic and multimodal workflows where speed and cost matter alongside visual reasoning.

Trinity-Large-Thinking is Arcee AI's open-source frontier reasoning model, a 399B-parameter sparse mixture-of-experts model (about 13B active per token) with a 512K-token context window, released under Apache 2.0. It generates explicit chain-of-thought reasoning traces and is built for complex, long-horizon agentic tasks and long-document analysis.

Grok 4.20 Multi-Agent is a variant of xAI's Grok 4.20 that runs several specialized agents on a shared backbone within a single API call to perform deep, multi-step research. It combines web search, code execution, and data synthesis tools to deliver comprehensive, well-sourced answers for research-heavy tasks.

Grok 4.20 is an xAI flagship model built for long-running agentic work and ambitious interactive and visual tasks, with a 500,000-token context window and text/image input. It targets high-intelligence coding, reasoning, and multi-step agent use cases.

Lyria 3 Pro is Google DeepMind's flagship music generation model, capable of producing full compositions up to three minutes long with structural awareness of intros, verses, choruses, and bridges. It supports multi-language vocals and genre versatility for professional-grade, prompt-driven music creation.

Lyria 3 (clip preview) is Google DeepMind's music generation model that produces high-fidelity stereo audio clips up to 30 seconds from text prompts. It is designed for rapid prototyping, social media assets, and short-form audio generation rather than full song composition.

Reka Edge is a small, efficient model from Reka AI designed for low-latency, on-device or resource-constrained deployments. It targets simple chat and task completion where speed and low cost outweigh the need for top-end capability.

MiniMax-M2.7 is MiniMax's model for software engineering, agent workflows, and professional document tasks, notable as a self-evolving model that ran autonomous reinforcement-learning optimization loops during training. It targets advanced coding and long-horizon agentic use cases.

GPT-5.4 Nano is the smallest, fastest model in the GPT-5.4 family, aimed at high-throughput, low-latency tasks like classification and simple extraction where cost efficiency is the priority.

GPT-5.4 Nano is the smallest, fastest model in the GPT-5.4 family, aimed at high-throughput, low-latency tasks. This is the batch variant for cheaper asynchronous processing.

GPT-5.4 Mini is the smaller, cheaper counterpart to GPT-5.4, suited to everyday chat, lightweight coding, and agentic tasks where speed and cost outweigh maximum reasoning depth.

GPT-5.4 Mini is the smaller, cheaper counterpart to GPT-5.4 for everyday chat, lightweight coding, and agentic tasks. This is the batch variant for cheaper asynchronous processing.

Mistral Small 4 (2603) is Mistral AI's March 2026 model, a 119B-parameter mixture-of-experts model with 6B active parameters that unifies instruct, reasoning, and coding with a toggleable reasoning mode and 256K context. It combines reasoning strength from Magistral, vision from Pixtral, and coding ability from Devstral in one efficient model.

This is the batch-processing variant of Mistral Small 4 (mistral-small-2603), Mistral AI's 119B-parameter mixture-of-experts model (6.5B active, Apache 2.0 license, 262K-token context) that unifies reasoning, vision, and agentic coding into one model. Positioned as an efficient mid-tier workhorse for complex analysis, coding, and multimodal tasks, the batch endpoint processes requests asynchronously at a lower price for high-volume, non-latency-sensitive workloads.

GLM-5-Turbo is a faster, lower-latency serving variant of Z.ai's flagship GLM-5 model, aimed at reducing cost and response time while retaining GLM-5's strengths in coding and long-horizon agentic workflows.

Nemotron 3 Super is a 120B mixture-of-experts model with 12B active parameters in NVIDIA's Nemotron 3 family, sitting between Nano and Ultra in scale. It offers stronger reasoning and agentic capability than Nemotron 3 Nano while remaining more efficient than Nemotron 3 Ultra.

Nemotron 3 Super is a 120B mixture-of-experts model with 12B active parameters in NVIDIA's Nemotron 3 family, offering stronger reasoning and agentic capability than Nemotron 3 Nano while staying more efficient than Nemotron 3 Ultra. This is the free-tier variant of the same model.

Seed 2.0 Lite is a cost-efficient, low-latency multimodal model from ByteDance Seed, delivering solid text, vision, and agentic tool-use capability. It is designed as a practical default for high-frequency production workloads.

Qwen3.5 9B is a small open-weight model from Alibaba's Qwen3.5 generation, suited for fast, low-cost general chat and simple coding tasks with the efficiency gains of the 3.5 series.

GPT-5.4 Pro is the extended-reasoning tier of GPT-5.4, prized for low hallucination rates on hard problems. It suits complex research, coding, and analysis where reliability matters more than speed.

GPT-5.4 Pro is the extended-reasoning tier of GPT-5.4 for complex research, coding, and analysis that need reliable, low-hallucination answers. This is the batch variant for cheaper asynchronous processing.

GPT-5.4 is OpenAI's March 2026 flagship model, released in Thinking and Pro variants known for low hallucination rates. It targets advanced reasoning, research, and coding tasks that need reliable, well-grounded answers.

GPT-5.4 is OpenAI's March 2026 flagship model known for low hallucination rates, suited to advanced reasoning, research, and coding. This is the batch variant for cheaper asynchronous processing.

Mercury 2 is Inception Labs' diffusion-based LLM that generates responses in parallel rather than token-by-token, delivering very high throughput (1000+ tokens/second) with a 128K context window. It offers tunable reasoning and native tool use, matching Claude Haiku/Gemini Flash-class quality at much faster speeds and lower cost, ideal for latency-sensitive agentic and coding tasks.

Gemini 3.1 Flash-Lite Preview is the preview release of Google's high-efficiency multimodal model, optimized for low-latency, high-volume workloads such as lightweight agentic workflows and simple data extraction. It offers the same responsiveness and cost focus as the GA Flash-Lite release.

Seed 2.0 Mini is ByteDance Seed's smallest 2.0-generation model, targeting latency-sensitive, high-concurrency, cost-sensitive scenarios with a 256K context window and four reasoning-effort modes. It delivers performance comparable to Seed 1.6 at lower cost.

Qwen3.5 35B-A3B is a sparse mixture-of-experts model from Alibaba's Qwen3.5 line (35B total, 3B active parameters), built for efficient general-purpose and agentic tasks at low inference cost.

Qwen3.5 27B is a mid-size dense open-weight model from Alibaba's Qwen3.5 generation, offering general-purpose chat, coding, and agentic capability with efficiency improvements over the Qwen3 series.

Qwen3.5 122B-A10B is one of Alibaba's Qwen3.5 open-weight mixture-of-experts models (122B total, 10B active), part of a generation positioned for the 'agentic AI era' with improved tool use and multi-step task handling over Qwen3.

Qwen3.5 Flash is Alibaba's fast, low-cost proprietary tier of the Qwen3.5 generation (this snapshot dated February 2026), designed for high-throughput, latency-sensitive general and agentic tasks.

Gemini 3.1 Pro Preview (custom tools) is a variant of Google's Gemini 3.1 Pro Preview configured for custom tool-calling setups, sharing the base model's strengths in agentic coding, long-horizon tool orchestration, structured planning, and multimodal analysis.

GPT-5.3-Codex is OpenAI's February 2026 Codex-native coding agent, pairing frontier coding performance with general reasoning for long-horizon, real-world engineering work. It supports steering the agent mid-task and is notable as a model OpenAI used to help build itself.

Aion-2.0 is a variant of DeepSeek V3.2 from AionLabs, tuned specifically for immersive roleplaying and collaborative storytelling. It is particularly strong at introducing tension, conflict, and mature or darker themes with nuance and depth.

Gemini 3.1 Pro Preview is Google's advanced reasoning model improving long-horizon stability and tool orchestration over Gemini 3 Pro, with a new medium thinking level to balance cost, speed, and performance. It excels at agentic coding, structured planning, multimodal analysis, and workflow automation for autonomous agents and enterprise tasks.

Gemini 3.1 Pro Preview is Google's advanced reasoning model improving long-horizon stability and tool orchestration, well suited to agentic coding, structured planning, multimodal analysis, and workflow automation. This is the batch variant for asynchronous, lower-cost processing.

Claude Sonnet 4.6 is a further refined mid-tier model, engineered for complex enterprise workflows, deep software engineering, and autonomous agentic execution with a 1M-token context window in beta. It scores strongly on coding (SWE-bench) and computer-use benchmarks while keeping Sonnet-tier pricing.

Batch-processing variant of Claude Sonnet 4.6, Anthropic's mid-tier model for enterprise coding and agentic execution. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

A February 2026 snapshot of Qwen3.5 Plus, Alibaba's proprietary mid-tier Qwen3.5 model, balancing cost and capability for general chat, reasoning, and agentic workloads.

Qwen3.5 397B-A17B is the largest and first-released model in Alibaba's Qwen3.5 mixture-of-experts generation, targeting flagship-level reasoning, coding, and agentic performance among open-weight models.

MiniMax-M2.5 is a further update in MiniMax's M2 series, achieving state-of-the-art results among MiniMax models in programming, tool calling, search, and office productivity tasks. It is aimed at real-world developer and knowledge-work productivity.

GLM-5 is Z.ai's flagship open-source foundation model, scaling to 744B total parameters (40B active) and trained on 28.5T tokens. It is engineered for complex systems design and long-horizon agent workflows, delivering production-grade performance on large-scale programming tasks for expert developers.

The extended chain-of-thought variant of Qwen3 Max, Alibaba's flagship model, built for the hardest reasoning, math, and multi-step agentic tasks where deliberate step-by-step thinking improves accuracy.

Claude Opus 4.6 is Anthropic's strongest model for coding and long-running professional tasks in this generation, with a 1M-token context window and 128K max output. It is designed for the most demanding software engineering and multi-step agentic workflows.

Batch-processing variant of Claude Opus 4.6, Anthropic's strongest coding-focused Opus model for long-running professional tasks. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

Qwen3 Coder Next is a next-generation iteration of Alibaba's Qwen3 Coder line, likely built on the more efficient Qwen3-Next architecture, aimed at agentic coding with improved efficiency and long-context repository handling.

OpenRouter Free is a meta-routing model from OpenRouter that automatically selects among available free-tier models to answer a request. It suits experimentation, low-stakes tasks, and cost-free testing rather than production workloads needing a specific model's guarantees.

Step 3.5 Flash is StepFun's flagship sparse MoE reasoning model (196B parameters, 11B active) built for fast, reliable execution of complex tasks. It excels at decomposing and planning multi-step problems and orchestrating tool calls, making it well suited for agentic and coding workflows.

Kimi K2.5 is Moonshot AI's January 2026 update to Kimi K2, a 1T-parameter mixture-of-experts model with 32B active parameters and multimodal vision-language understanding, offering both instant and thinking modes. It is built for advanced agentic and conversational workflows.

Solar Pro 3 is Upstage's efficient large language model, building on Solar Pro 2 with Upstage's SnapPO reinforcement learning framework to sharpen step-by-step reasoning in math, code, and agentic tasks. It roughly doubles agentic benchmark performance over its predecessor while keeping the same API and serving footprint, with strong multilingual (notably Korean) support.

MiniMax-M2-her is a dialogue-first model built for immersive roleplay, character-driven chat, and expressive multi-turn conversation, distilled from years of MiniMax's role-play optimization work. It powers MiniMax's Talkie companion app and is suited to storytelling and companion-style interactions rather than coding or agentic tasks.

Palmyra X5 is Writer's most advanced enterprise foundation model, built for building and scaling AI agents with a 1M-token context window and a hybrid attention architecture. It offers advanced reasoning, multi-step tool-calling, built-in RAG, and multilingual support tailored for enterprise workflows.

GPT-Audio is OpenAI's model for natural speech input and output, generating expressive spoken responses directly from audio prompts. It is designed for voice assistants, spoken dialogue apps, and other audio-native interactions.

GPT-Audio Mini is a smaller, cheaper version of OpenAI's GPT-Audio model for speech input and output, suited to cost-sensitive or high-volume voice assistant and spoken dialogue applications.

GLM-4.7-Flash is a 30B-class SOTA model from Z.ai that balances performance and efficiency, offering a faster, cheaper alternative to the full GLM-4.7 model. It targets general reasoning and coding tasks where speed and cost matter more than maximum capability.

GPT-5.2-Codex is the coding-specialized variant of GPT-5.2, built for agentic software engineering across large codebases. It benefits from GPT-5.2's large context window for repository-scale refactors and long-horizon coding tasks.

Seed 1.6 Flash is a faster, lower-latency variant of ByteDance Seed's Seed 1.6 multimodal model, trading some depth for speed and cost efficiency. It suits high-throughput text and vision tasks where responsiveness matters most.

Seed 1.6 is ByteDance Seed's general-purpose multimodal model with adaptive deep-thinking capability and a 256K-token context window. It is designed as a versatile foundation model for text, vision, and reasoning tasks.

MiniMax-M2.1 is an update to MiniMax's M2 coding and agentic model line, focused on polyglot programming mastery and precision code refactoring. It targets developers needing strong multi-language coding assistance and reliable agentic tool use.

GLM-4.7 is Z.ai's open-source model generation released after GLM-4.6, continuing incremental improvements in reasoning and coding. It serves as the base for the GLM-4.7-Flash variant that balances performance and efficiency at a smaller 30B class size.

Gemini 3 Flash Preview is Google's fast workhorse model in the Gemini 3 line, tuned for agentic workflows, coding, and complex multi-step reasoning at low latency. It became the default model in the Gemini app for everyday responsive tasks.

Gemini 3 Flash Preview is Google's fast workhorse model in the Gemini 3 line, tuned for agentic workflows, coding, and complex multi-step reasoning at low latency. This is the batch variant for asynchronous, lower-cost processing.

Nemotron 3 Nano is NVIDIA's compute-efficient open model in the Nemotron 3 family, a 30B mixture-of-experts model with 3B active parameters. It is optimized for high-throughput, low-cost agentic and multi-agent workloads.

GPT-5.2 Chat is the fast, conversational (instant) variant of GPT-5.2, tuned for responsive everyday chat rather than deep extended reasoning. It suits general assistant use cases where low latency matters most.

GPT-5.2 Pro is the extended-reasoning tier of GPT-5.2, trading latency for maximum accuracy on hard problems. It is aimed at complex research, coding, and analytical work with OpenAI's largest available context window.

GPT-5.2 Pro is the extended-reasoning tier of GPT-5.2 for complex research, coding, and analytical work. This is the batch variant for cheaper asynchronous processing.

GPT-5.2 is OpenAI's December 2025 flagship release, offered in instant, thinking, and pro modes with a 400K-token context window. It is designed for enterprise-scale reasoning, document and codebase analysis, and complex agentic workflows.

GPT-5.2 is OpenAI's December 2025 flagship release with a 400K-token context window, suited to enterprise-scale reasoning, document analysis, and agentic workflows. This is the batch variant for cheaper asynchronous processing.

Devstral 2 2512 is Mistral AI's open-weight agentic coding model (123B-parameter dense transformer, 256K-token context), released December 2025 and built for tool use that explores codebases, edits multiple files, and powers software-engineering agents. It scores 72.2% on SWE-Bench Verified and 61.3% on SWE-Bench Multilingual, positioning it as a top open-source model for project-level coding and long-horizon agent workflows.

Relace Search is a companion model to Relace's code-apply system, built for fast retrieval of relevant code context and locations within a codebase. It is designed for AI coding agents that need quick, accurate code search before generating or applying edits.

GLM-4.6V is Z.ai's vision-language model that treats images, video, and tools as first-class agent inputs, extending training context to 128K tokens with native multimodal function calling. It is designed for agentic multimodal workflows such as screenshot- and document-driven tool use.

GPT-5.1-Codex-Max is OpenAI's frontier agentic coding model, using a compaction technique to coherently work across millions of tokens in a single task. It targets project-scale refactors, extended debugging sessions, and multi-hour autonomous coding loops.

Amazon Nova 2 Lite is a fast, cost-effective multimodal reasoning model from AWS that handles text, image, video, and document input with a 1M-token context window. It supports optional extended thinking with adjustable reasoning budgets plus built-in web grounding and code interpreter tools, aimed at everyday production workloads that need quick, affordable responses.

Ministral 3 14B is Mistral AI's December 2025 edge-focused model, offered in base, instruct, and reasoning variants with a 256K context window. It targets on-device and edge deployment with function calling and multilingual support, offering the strongest quality of the Ministral 3 sizes.

Ministral 3 8B is the mid-sized model in Mistral AI's December 2025 Ministral 3 edge family, balancing capability and efficiency for on-device deployment. It supports a 256K context window, function calling, and multilingual chat.

This is the batch-processing variant of Ministral 3 8B (ministral-8b-2512), Mistral AI's dense 8.4B-parameter model plus a 0.4B vision encoder (Apache 2.0 license, 262K-token context, multilingual and multimodal), released December 2025. Designed as a small, efficient edge-capable model that fits in 12GB of VRAM, it suits lightweight and on-device deployments, and the batch endpoint adds a further cost reduction for high-volume asynchronous processing.

Ministral 3 3B is the smallest model in Mistral AI's December 2025 Ministral 3 edge family, designed for lightweight on-device deployment. It trades some capability for a tiny footprint, suiting simple chat, classification, and function-calling tasks on constrained hardware.

Mistral Large 3 (2512) is Mistral AI's December 2025 flagship model, a sparse mixture-of-experts architecture with 41B active of 675B total parameters, released under Apache 2.0. It is Mistral's most capable model to date for advanced reasoning, coding, and multilingual enterprise workloads.

This is the batch-processing variant of Mistral Large 3 (mistral-large-2512), Mistral AI's 675B-parameter sparse mixture-of-experts flagship (41B active, Apache 2.0 license, 262K-token context) released December 2025. It targets complex reasoning, agentic, and long-context workloads, and the batch endpoint offers the same capability at a reduced price via asynchronous processing, ideal for large-scale, non-real-time jobs.

DeepSeek V3.2 is DeepSeek's V3-generation model upgrade underlying both the deepseek-chat (non-thinking) and deepseek-reasoner (thinking) modes, improving efficiency and reasoning quality over V3.1. It is designed as a strong, low-cost general-purpose and reasoning model.

Claude Opus 4.5 is a further refinement of Anthropic's Opus flagship line, with improved coding, reasoning, and long-horizon agentic capabilities over Opus 4.1. It targets the most demanding professional software engineering and analysis tasks.

Batch-processing variant of Claude Opus 4.5, Anthropic's refined Opus flagship for demanding coding and reasoning work. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

GPT-5.1 is OpenAI's late-2025 successor to GPT-5, improving conversational warmth, instruction following, and reasoning quality. It suits general-purpose chat, coding, and agentic workflows that need more natural, steerable responses.

GPT-5.1 is OpenAI's late-2025 successor to GPT-5, improving conversational warmth, instruction following, and reasoning quality for chat, coding, and agentic workflows. This is the batch variant for cheaper asynchronous processing.

GPT-5.1-Codex is a coding-focused variant of GPT-5.1 optimized for long-running, agentic software engineering tasks in Codex or Codex-like harnesses. It is well suited to multi-step refactors, debugging, and repository-scale coding work.

GPT-5.1-Codex-Mini is a smaller, cheaper version of GPT-5.1-Codex for agentic coding tasks. It suits lighter coding workflows where speed and cost matter more than the largest context or maximum autonomy.

Kimi K2 Thinking is a reasoning-focused variant of Moonshot AI's Kimi K2 that produces extended chain-of-thought before answering. It is designed for complex multi-step reasoning, math, and agentic planning tasks that benefit from deliberate thinking.

Amazon Nova Premier is the most capable model in AWS's original Nova lineup, built for complex reasoning, multimodal understanding, and use as a teacher model for distillation. It targets demanding enterprise workloads requiring the highest accuracy in the Nova family.

Sonar Pro Search is a variant of Perplexity's Sonar Pro tuned for heavier, broader web search and retrieval, pulling in more sources per query. It suits research-heavy tasks where thorough web coverage matters more than raw response speed.

Voxtral Small is Mistral AI's 24B audio-native model for speech understanding, transcription, and voice-driven instruction-following. It is designed for tasks combining spoken audio input with text reasoning, such as voice assistants and audio Q&A.

GPT-OSS-Safeguard-20B is an open-weight variant of GPT-OSS-20B fine-tuned for safety classification and content moderation, letting developers apply and customize policy-based filtering against custom safety guidelines.

MiniMax-M2 is MiniMax's model built for coding and agentic workflows, offering strong tool-use, multi-step task execution, and software engineering performance. It is designed for developers building autonomous coding agents and complex task automation.

Qwen3 VL 32B Instruct is a dense multimodal model in Alibaba's Qwen3 VL family, handling image understanding, OCR, and visual reasoning alongside general chat and coding ability.

Granite 4.0 H-Micro is IBM's small, efficient dense language model using a hybrid architecture, optimized for low-latency, cost-efficient workloads. It significantly outperforms earlier same-size Granite models, making it suitable for lightweight chat and instruction-following tasks.

Claude Haiku 4.5 is Anthropic's fast, affordable small model in the Claude 4 generation, balancing strong instruction-following and coding ability with low latency. It is designed for high-throughput agentic and chat applications where speed and cost efficiency are priorities.

Batch-processing variant of Claude Haiku 4.5, Anthropic's fast and affordable Claude 4 small model. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

The extended-reasoning variant of Qwen3 VL 8B, adding step-by-step chain-of-thought to the small multimodal model for improved accuracy on visual reasoning tasks while staying lightweight.

Qwen3 VL 8B Instruct is a small, fast multimodal model from Alibaba's Qwen3 VL line, suited for lightweight image understanding and vision-plus-text tasks where speed and cost matter most.

The extended-reasoning variant of Qwen3 VL 30B-A3B, providing step-by-step reasoning over visual and text inputs for more complex multimodal analysis while remaining cheaper than the largest Qwen3 VL models.

Qwen3 VL 30B-A3B Instruct is a smaller mixture-of-experts multimodal model from Alibaba's Qwen3 VL family, offering solid image and document understanding at lower inference cost than the 235B version.

GPT-5 Pro is an extended-reasoning tier of GPT-5 that trades latency and cost for higher accuracy on the hardest problems. It is aimed at difficult research, coding, and analytical tasks that benefit from maximum reasoning effort.

GPT-5 Pro is an extended-reasoning tier of GPT-5 for the hardest research, coding, and analytical tasks. This is the batch variant for cheaper asynchronous processing.

GLM-4.6 is Z.ai's open-source flagship model, an incremental upgrade over GLM-4.5 with improved reasoning, coding, and agentic capabilities. It continues the GLM-4.5 architecture lineage while pushing higher benchmark performance for complex, production-grade tasks.

Claude Sonnet 4.5 is an upgraded mid-tier model in the Claude 4 line, improving coding, computer use, and long-context reasoning over Sonnet 4 while keeping Sonnet-level pricing. It is well suited for enterprise workflows, agentic coding, and general-purpose assistant tasks.

Batch-processing variant of Claude Sonnet 4.5, Anthropic's upgraded mid-tier Claude 4 model. Same capabilities as the standard model, offered at lower cost via asynchronous batch processing.

DeepSeek V3.2-exp is an experimental preview checkpoint of DeepSeek V3.2, used to test architecture or training improvements ahead of the stable V3.2 release. It targets the same general reasoning, coding, and chat use cases as V3.2 with less production-hardening.

Cydonia 24B v4.1 is a community fine-tune from TheDrummer built on a Mistral Small 24B base, tuned for creative writing and roleplay. It favors expressive, character-driven prose over the more clinical tone of general-purpose assistant models.

Relace Apply 3 is a specialized code-patching model that merges AI-suggested edits directly into existing source files at very high speed (thousands of tokens per second). It is designed to be paired with a larger generation model as a fast, cheap 'instant apply' layer, reducing latency and token cost in AI coding tools.

The extended-reasoning variant of Qwen3 VL 235B-A22B, Alibaba's large multimodal model, which reasons step by step over visual and textual inputs for complex visual reasoning and analysis tasks.

Qwen3 VL 235B-A22B Instruct is Alibaba's large-scale multimodal MoE model, combining Qwen3's language capability with vision understanding for image, document, and video tasks. It suits demanding multimodal applications needing near-flagship quality.

Qwen3 Max is Alibaba's flagship proprietary Qwen3 model, aimed at top-tier reasoning, coding, and agentic performance to compete with other frontier models like GPT and Claude. It is the highest-capability option in the Qwen3 lineup.

Qwen3 Coder Plus is a higher-capability tier of Alibaba's Qwen3 Coder family, designed for more demanding agentic coding and software engineering tasks that need stronger reasoning over large codebases than the flash/base variants.

DeepSeek V3.1 Terminus is a refined checkpoint of DeepSeek V3.1, tuned for improved stability, agentic tool use, and coding performance. It is positioned as a polished general-purpose and agentic workhorse ahead of the V3.2/V4 generations.

Qwen3 Coder Flash is a faster, lower-cost variant of Alibaba's Qwen3 Coder model, trading some capability for speed. It suits latency-sensitive coding assistants and high-volume code-completion workloads.

The extended-reasoning variant of Qwen3-Next 80B-A3B, Alibaba's efficient next-generation MoE model, producing explicit chain-of-thought for harder reasoning and agentic tasks while retaining the architecture's low inference cost.