AI Models Catalog

Zhipu's August 14, 2026 coding and agentic model, built on the GLM-5.2 base with further post-training. Z.ai reports Terminal-Bench 3.0 of 28.3 and DeepSWE v1.1 of 66.9. Published weights use the custom GLM-5.3 license; hosted access is available through Z.ai and the GLM Coding Plan. Text-only input, a documented 1M-token context window and up to 128K output, with reasoning always enabled. Useful for coding teams that need long-running agents or self-hosted deployment under the model license.

Alibaba's August 14, 2026 dense 27B vision-language model, with Apache 2.0 weights and optional thinking. It accepts text, images and video. Published weights have a native 262,144-token context, extensible to 1M with YaRN; QwenCloud provides a hosted 1M-context endpoint at $0.50/$3 per MTok (input/output). Hosted limits and prices are separate from local deployment requirements.

Google's Flash model for coding and agent workflows, released August 13, 2026. At launch, Google reported DeepSWE v1.1 65.3%, FrontierCode 1.1 Main 43.6%, WebDev Arena 1588 Elo and AutomationBench 30.4%. It accepts text, images, audio, video and PDFs, with a 1,048,576-token input limit and a 65,536-token text-output limit. Thinking levels are low, medium and high. Standard Gemini API rates are $0.75/$3.75 per MTok (input/output) through December 31, 2026, then $1.50/$7.50 from January 1, 2027.

Alibaba's text-only, post-trained model underlying Qwen3.8-Max, released August 12, 2026. The 2.4T-parameter MoE activates 95B parameters and combines Gated DeltaNet with Gated Attention. Native context is 262,144 tokens, extensible to about 1M. Thinking is always enabled, with low, medium, and xhigh effort. Published weights use the custom Qwen3.8-Max License; hosted Max adds vision, non-thinking mode, 1M default context, and built-in tools.

xAI's August 12, 2026 model for coding, agentic tasks and knowledge work. It builds on Grok 4.5 with supplemental training, regenerated supervised fine-tuning trajectories and agentic reinforcement learning. At launch, xAI reported DeepSWE v1.1 65.9% and APEX-Agents 57.5% at high reasoning effort. It accepts text and images, produces text and has a 500K-token context window, with low/medium/high/xhigh reasoning effort. Standard xAI API input/output rates are $2/$6 per MTok (input/output) below 200K prompt tokens and $4/$12 at or above 200K; cached input is $0.50/$1 respectively. Microsoft Foundry availability was announced August 26, 2026.

Meta's 30B dense model for local agents, released August 10, 2026 with downloadable Apache 2.0 weights. Accepts text and images and produces text, with a 128K-token context window. Quantized language-model weights fit under 20GB; Meta targets 24GB or 32GB total memory for the weights, working memory, perception encoder and speculative-decoding drafter. Supports tool use, coding and multi-step agent workflows on consumer hardware.

Meta's August 5, 2026 coding-focused update to Muse Spark 1.1, released in Muse Code and Meta Model API. Co-trained with Muse Code for code generation, debugging, codebase understanding and long-horizon developer workflows. Supports image and video reasoning and audiovisual workflows. Meta launched the later Muse Spark 1.3 generation on September 2, 2026.

Alibaba's hosted Qwen3.8-Max accepts text, images and video within a 1M-token context window. QwenCloud lists $2/$6 per MTok (input/output). Alibaba Cloud added the dated qwen3.8-max-0902 snapshot on September 2, 2026, also named qwen3.8-max-2026-09-02. The related Qwen3.8-2.4T-A95B weights are a text-only post-trained model; their license and local context limits should not be conflated with the hosted service.

MiniMax's July 31, 2026 Hailuo 3.0 video model family generates 4-15 second clips at up to 2K / 24 FPS with stereo audio and multi-shot support in the complete system. Published H3-Base weights generate 768p locally under the MiniMax community license; the full 2K workflow also requires hosted processing and H3-Regenerate-2K. Omni Reference accepts text, image, video and audio inputs, with up to 12 mixed reference files.

Anthropic's July 24, 2026 model for complex coding and enterprise work, with adjustable effort and thinking enabled by default. Provides a 1M-token context window and up to 128K output at $5 input and $25 output per million tokens, unchanged from Opus 4.8. Optional Fast Mode is a research preview on the Claude API with access restrictions: up to 2.5x output-token throughput at $10/$50 per million tokens.

Google's July 21, 2026 Flash model for coding, agentic tasks and multimodal analysis, superseded as the workhorse Flash by Gemini 3.7 Flash on August 13, 2026. Accepts text, images, video, audio and PDFs, with 1M input tokens and up to 65K text output. Standard API pricing is $0.75 input and $3.75 output per million tokens through December 31, 2026, then $1.50/$7.50 from January 1, 2027.

Google's July 21, 2026 low-latency model for subagent tasks and high-volume document processing. Accepts text, images, video, audio and PDFs, with 1M input tokens and up to 65K text output. Standard API rates are $0.30 input and $2.50 output per million tokens; batch rates are $0.15/$1.25. Standard cached input costs $0.03 per million tokens, with cache storage billed separately.

Moonshot's Kimi K3 is a native vision model with 2.8T total and 104B active parameters and a 1,048,576-token context. Published weights use the Kimi K3 License. The first-party API supports text, image and video inputs; thinking is always enabled with low, high and max effort settings. It is available for coding and knowledge-work agents through Moonshot's API and self-hosted weights.

OpenAI's July 9, 2026 GPT-5.6 model for complex professional work. Supports text and image input, text output and adjustable reasoning effort, with a 1.05M-token context window and up to 128K output. Promotional API rates are $4 input and $20 output per million tokens through at least November 21, 2026; cached input is $0.40. Above 272K input tokens, the full request costs 2x input and 1.5x output rates. Cache writes cost 1.25x the uncached input rate.

OpenAI's July 9, 2026 GPT-5.6 model balancing intelligence and cost. Supports text and image input, text output and adjustable reasoning effort, with a 1.05M-token context window and up to 128K output. API prices fell 20% on July 30 to $2 input and $12 output per million tokens; cached input is $0.20. Above 272K input tokens, the full request costs 2x input and 1.5x output rates. Cache writes cost 1.25x the uncached input rate.

OpenAI's July 9, 2026 GPT-5.6 model for cost-sensitive, high-volume workloads. Supports text and image input, text output and adjustable reasoning effort, with a 1.05M-token context window and up to 128K output. API prices fell 80% on July 30 to $0.20 input and $1.20 output per million tokens; cached input is $0.02. Above 272K input tokens, the full request costs 2x input and 1.5x output rates. Cache writes cost 1.25x the uncached input rate.

xAI's coding, agentic and knowledge-work model, announced publicly on July 16, 2026; its API release notes record availability on July 8. It accepts text and images, produces text and supports function calling, structured outputs and reasoning within a 500K-token context window. Standard xAI API input/output rates are $2/$6 per MTok (input/output) below 200K prompt tokens and $4/$12 at or above 200K; cached input is $0.30/$0.60 respectively.

Alibaba's May 2026 text-only Qwen3.7-Max remains listed on QwenCloud with a 1M-token context and up to 131K output. It supports thinking and tool use for coding and knowledge work. QwenCloud currently lists $2.50/$7.50 per MTok (input/output). Qwen3.8-Max is the newer hosted generation.

Alibaba's June 2026 Qwen3.7-Plus is a hosted model accepting text, images and video with a 1M-token context. Its preserve_thinking option retains reasoning across tool calls. Current provider rates depend on input length and promotional discounts, so a single price ratio against Max does not describe all requests.

Anthropic's June 9, 2026 restricted-access counterpart to Fable 5, available by invitation through Project Glasswing. It shares Fable 5's capabilities but does not include the same safety classifiers, so behavior is not identical. It supports text and image input, a 1M-token context window, up to 128K output and always-on thinking. Mythos 5.1 was released September 1; Mythos 5 remains listed as active.

Anthropic's June 9, 2026 Fable model remains active as a prior-generation option after Fable 5.1 launched September 1. It accepts text and images, provides a 1M-token context window and up to 128K output, and uses always-on thinking. Standard Claude API prices are $10/$50 per MTok (input/output), with $1 cached input. Fable 5 includes safety classifiers that can refuse requests; its restricted Mythos 5 counterpart has the same capabilities without those classifiers.

Moonshot's Kimi K2.7 Code is a native multimodal coding model with 1T total and 32B active parameters and a 256K-token context. It accepts text, images and video; thinking is mandatory. Moonshot reports about 30% fewer thinking tokens than K2.6 on coding and agentic evaluations. Published weights use a modified MIT license, and the first-party API is OpenAI-compatible.

Microsoft's first fully in-house reasoning model, unveiled at Build 2026 (June 2) — a sparse MoE with ~35B active of ~1T total params, trained without OpenAI distillation using data Microsoft describes as appropriately licensed. AIME 2025 97.0%, AIME 2026 94.5%, SWE-Bench Pro competitive with Claude Opus 4.6, and preferred over Sonnet 4.6 in blind human evals across 1,276 tasks. 256K context; public preview on Microsoft Foundry since August 12, 2026 with third-party availability via Fireworks AI, Baseten, and OpenRouter.

Anthropic's June 30, 2026 Sonnet model for coding and agent workflows, with text and image input, a 1M-token context window and up to 128K output. Thinking is enabled by default with adjustable effort. Standard Claude API pricing is $2/$10 per MTok (input/output); the previously announced September 1 price increase was canceled. Anthropic's launch comparisons with Opus 4.8 depend on the task and effort setting.

Anthropic's May 28, 2026 Opus model remains active as a legacy option after Opus 5 launched July 24. It supports a 1M-token context window, up to 128K output and adaptive thinking at standard Claude API rates of $5/$25 per MTok (input/output). In Anthropic's launch evaluation, it was about four times less likely than Opus 4.7 to leave flaws in code it had written unremarked. Dynamic subagent workflows launched separately as a Claude Code research preview.

xAI's Grok 4.3 accepts text and images and produces text within a 1M-token context window. Its reasoning effort can be set to none, low, medium or high, and it supports function calling and structured outputs. Standard xAI API input/output rates are $1.25/$2.50 per MTok (input/output) below 200K prompt tokens and $2.50/$5 at or above 200K. Current-event retrieval requires enabled search tools.

Mistral's April 28, 2026 model (v26.04), combining instruction-following, reasoning, and coding in 128B dense parameters. It supports vision, function calling, configurable reasoning, and a 256K-token context. Published weights use a Modified MIT license. Mistral lists standard provider API pricing of $1.50/$7.50 per MTok for input/output. Its May 22 product announcement made the model the default in Vibe CLI and Le Chat.

DeepSeek-V4-Pro-0813 is the August 13, 2026 text-only V4 Pro release, with 1.6T total and 49B active parameters, a 1M-token context and MIT weights. As of September 10, the deepseek-v4-pro API still identifies this build. DeepSeek has announced that from September 14, 2026 at 12:00 Beijing time (04:00 UTC), this slug will temporarily route to DeepSeek-V4.1-Flash at Flash pricing until a future V4.1 Pro release.

DeepSeek-V4-Flash-0731 is the July 31, 2026 text-only V4 Flash release, with 284B total and 13B active parameters, a 1M-token context and MIT weights. DeepSeek retired its first-party V4 Flash service on September 10, 2026. The legacy deepseek-v4-flash slug now temporarily routes to the newer, vision-capable DeepSeek-V4.1-Flash. These historical weights retain their own architecture and modalities.

OpenAI's GPT-5.5 launched in ChatGPT and Codex on April 23, 2026, with API access announced April 24. It remains listed in the API as a prior-generation model with text and image input, text output, a 1.05M-token context window and up to 128K output. Standard API rates are $5/$30 per MTok (input/output), with $0.50 cached input; sessions exceeding 272K input tokens use 2x input and 1.5x output rates.

A GPT-5.5 reasoning variant that uses additional compute for complex questions. Supports medium, high and xhigh reasoning effort, text and image input, and text output. Available through the Responses API, including Batch requests, with a 1M-token context window and up to 128K output tokens. Some requests can take several minutes; background mode supports longer tasks.

The GPT-5.5 reasoning option introduced in ChatGPT on April 23, 2026 for complex professional work. Reasoning summaries can help explain an answer, but do not expose the model's raw hidden reasoning. This is a historical ChatGPT variant; availability and usage limits depend on the product and plan.

OpenAI's May 5, 2026 ChatGPT conversation model, retained here as a historical variant after GPT-5.6 Luna became the Free default. In OpenAI's internal launch evaluation of high-stakes prompts, it produced 52.5% fewer hallucinated claims than GPT-5.3 Instant. Its original API access used the moving chat-latest alias, which no longer establishes a fixed GPT-5.5 Instant identity.

Moonshot's Kimi K2.6 remains available alongside K3. Its 1T-parameter MoE activates 32B parameters and supports a 256K-token context, text, images and video, with thinking or instant responses. Moonshot reports orchestration of up to 300 sub-agents over 4,000 coordinated steps and a SWE-Bench Pro score of 58.6 under its published evaluation setup.

Alibaba's legacy Qwen3.6-Max-Preview is a text-only hosted model with a 256K-token context. Alibaba Cloud Model Studio has scheduled this endpoint for retirement on October 10, 2026 at 00:00 UTC+8, subject to the actual rollout time.

Anthropic's April 16, 2026 Opus model remains active as a legacy option, with a 1M-token context window, up to 128K output and text and image input. Its launch emphasized software engineering and higher-resolution vision. Task budgets were introduced in public beta to guide token spending during agent loops. Current documentation recommends Opus 5 for new Opus workloads.

Alibaba's April 16, 2026 Apache 2.0 vision-language model uses a 35B-parameter MoE with 3B active parameters. It supports text, image and video inputs and local deployment. Published weights support 262,144 tokens natively, extensible to 1,010,000 with YaRN.

Google DeepMind's open-weight Gemma 4 family, announced April 2, 2026 under Apache 2.0. The family now includes E2B, E4B, 12B, 26B A4B and 31B; 12B Unified was added June 3. E2B/E4B support 128K context, while 12B/26B A4B/31B support 256K. All accept text and images and produce text; audio input is supported by E2B, E4B and 12B. The family supports configurable thinking and function calling for local and server deployments.

Alibaba introduced Qwen3.6-Plus on April 2, 2026 for coding agents and visual workflows. The model has a 1M-token context and remains listed among Model Studio's legacy models. Its original announcement describes repository-level engineering and interaction with visual environments.

Z.ai's GLM-5.2 is a text-only model for coding and agent workflows, with a 1M-token context and up to 128K output. It remains available in the first-party API after the GLM-5.3 release. Published weights use the MIT license. The documented context is five times GLM-5.1's 200K limit.

Z.ai lists the GLM-5.1 release on April 7, 2026. This text-only coding and agent model has a 200K-token context and up to 128K output. Published weights use the MIT license, and the model remains listed in first-party API pricing alongside newer GLM releases.

MiniMax introduced M3 on June 1, 2026 with MiniMax Sparse Attention, native text, image and video input, and a 1M-token context. Published weights use the MiniMax community license. The standard provider API costs $0.30/$1.20 per MTok (input/output) for requests with at most 512K input tokens; both rates double above that threshold. Optional priority service costs 1.5 times the corresponding standard rate. MiniMax labels these standard rates a permanent 50% discount.

MiniMax's March 18, 2026 model for coding agents and office workflows, now superseded by M3. Supports multi-agent collaboration and editing Word, Excel and PowerPoint files through agent tools. MiniMax reports that an internal M2.7 research agent handled 30-50% of its reinforcement-learning team's workflow during development, with researchers guiding key decisions.

Mistral's July 30, 2025 model for low-latency code completion and fill-in-the-middle workflows. Current API documentation lists a 128K-token context and supports code generation, function calling and structured outputs. Enterprise deployment options include cloud, private cloud and on-premises infrastructure.

Mistral's March 16, 2026 unified model: merges Magistral (reasoning), Pixtral (vision), and Devstral (agentic coding) into a single 119B-total MoE with 6.5B active per token. 256K context, Apache 2.0, configurable reasoning effort.

Microsoft's 14B dense reasoning model, released April 30, 2025 under the MIT license. It accepts and produces text with a 32K-token context. Supervised fine-tuning emphasizes mathematical reasoning, science and coding. Designed for reasoning research and applications with constrained memory or compute, with downloadable weights for local deployment.

Microsoft's reinforcement-learning variant of Phi-4-reasoning, released April 30, 2025 with MIT-licensed weights. It retains the 14B dense architecture, text input and output, and 32K-token context. Microsoft reports longer reasoning responses and higher latency than the base reasoning variant, making it useful when additional reasoning time is acceptable.

OpenAI's December 11, 2025 GPT-5.2 model remains listed as a previous-generation API model. It accepts text and images and produces text, with a 400K-token context window and up to 128K output. Reasoning effort supports none, low, medium, high and xhigh. OpenAI's current model page recommends GPT-6 Astra for most new API workloads.

Anthropic's November 24, 2025 Opus model remains active as a legacy option. It accepts text and images and produces text, with a 200K-token context window, up to 64K output and extended thinking. Standard Claude API rates are $5/$25 per MTok (input/output). Current documentation recommends Opus 5 for new Opus workloads.