THE ARTICLE · 5 MIN
The companies that build large language models are often called AI labs. Some sell access to closed models through apps and an API. Others also publish the model’s weights so anyone can download and run it.
This page lists the main labs and each one’s current top model, as read from the labs’ own documentation on 25 September 2026. Model line-ups change every few weeks, so treat this as a dated snapshot and check the lab’s model page for anything newer.
The table
Labs are listed alphabetically. “Context window” is the figure each lab publishes for how many tokens a model can handle at once. Labs do not all count it the same way: for some it includes the model’s reply as well as your input.
| Lab | Current top model | Released | Weights | Context window |
|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-Max; weights published as Qwen3.8-2.4T-A95B | API: 3 August 2026 | Downloadable, custom licence | 262,144 natively, extensible to 1,010,000 (weights) |
| Anthropic | Claude Fable 5.1; Claude Opus 5.5 | 1 September 2026; 22 September 2026 | Closed | 1M tokens |
| DeepSeek | DeepSeek-V4-Pro; DeepSeek-V4.1-Flash | 13 August 2026 (V4-Pro general release); 10 September 2026 | Downloadable, MIT | 1M tokens |
| Google DeepMind | Gemini 3.1 Pro (preview); Gemini 3.8 Flash | 19 February 2026; 2 September 2026 (general availability) | Closed | 1,048,576 |
| Meta | Muse Spark 1.3 | 2 September 2026 (Muse Spark first announced 8 April 2026) | Closed | 1M tokens |
| Mistral AI | Mistral Medium 3.5 | April 2026 | Downloadable, custom licence | 256k |
| Moonshot AI | Kimi K3 | July 2026 | Downloadable, custom licence | 1,048,576 |
| OpenAI | GPT-6 Astra | 3 September 2026 | Closed | 1,050,000 |
| xAI (SpaceXAI) | Grok 4.7 | API: 21 September 2026 | Closed | 500k |
| Z.ai (Zhipu) | GLM-5.3 | 18 August 2026 | Downloadable, custom licence | 1M tokens |
What the labs say about their own models
These are the labs’ own descriptions, not independent assessments.
- OpenAI: its model index says “If you’re not sure where to start, use GPT-6 Astra”. Its system card is headed “Published September 3, 2026 GPT-6 Astra System Card”.
- Anthropic: its documentation says to “start with Claude Opus 5.5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short”.
- Google: it describes Gemini 3.8 Flash as “our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows”. Gemini 3.1 Pro is still listed as a preview, with “No shutdown date announced”.
- xAI: SpaceX announced on 2 February 2026 that it had acquired xAI, and xAI’s own pages now use the name SpaceXAI. Its models page calls Grok 4.7 “Our flagship model for code and everything else”.
- Meta: “Today we are announcing Muse Spark, the first in a new series of large language models built by Meta Superintelligence Labs.” At launch it said: “It will be available in private preview via API to select partners, and we hope to open-source future versions of the model.” Its developer page now describes the API as “Public preview. Now available with expanded global access.” Announcing version 1.3, Meta listed “the Muse Spark open weights release” on its roadmap.
- DeepSeek: its model card says the 0813 release is the official V4-Pro, “superseding the preview version, with greatly enhanced agentic capabilities”.
- Alibaba: its model card says “For the first time, Qwen3.8 brings a Qwen-Max-class model to open release”, with “2.4T in total and 95B activated” parameters.
- Mistral AI: its documentation says the model was “Released as open weights under a Modified MIT license.”
- Moonshot AI: “Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date.”
- Z.ai: “GLM-5.3 currently supports text-only inputs, with a 1M-token context window and a maximum output length of 128K tokens.”
Closed or downloadable?
A closed model’s weights are not published: you use it through an app or an API, and some closed models are also sold through other companies’ cloud platforms, such as Amazon Bedrock or Microsoft Foundry. A model with downloadable weights can be run on hardware you control — but “downloadable” is not the same as “open-source”, and the licence decides what you may do with it. Several of the downloadable models above use custom licences with extra conditions for large companies. See Open-weight is not open-source for what those licences say. Some labs publish both kinds: Meta’s Muse Glimmer model (30B parameters), for example, is released with open weights under Apache 2.0.
Models get retired
Labs publish schedules for switching old models off, and those dates matter if you build on a model:
- Anthropic’s deprecation page says of Claude Opus 4 and Claude Sonnet 4: “These models were retired June 15, 2026.”
- OpenAI says it notified developers of older GPT-5 and o3 snapshots of “their deprecation and removal from the API on December 11, 2026”.
- Google’s deprecation page notes: “The shutdown dates listed in the table indicate the earliest possible dates”.
- Plans can also change. DeepSeek’s change log says: “we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026”.
How to check what is current
Each lab keeps an official model page. The ones used for this table were:
- OpenAI: developers.openai.com/api/docs/models
- Anthropic: platform.claude.com/docs/en/about-claude/models/overview
- Google: ai.google.dev/gemini-api/docs/models
- xAI: docs.x.ai/docs/models
- Meta: developer.meta.com/ai/models/muse-spark
- DeepSeek: api-docs.deepseek.com
- Alibaba Qwen: huggingface.co/Qwen and docs.qwencloud.com
- Mistral AI: docs.mistral.ai/models
- Moonshot AI: huggingface.co/moonshotai
- Z.ai: docs.z.ai
Our reading: the table tells you who makes what, not which model is best. Labs’ descriptions of their own models are not independent assessments, and published scores depend heavily on how models are tested — see AI benchmarks explained. The practical test is how a model does on your own tasks.
Sources
- OpenAI: API models index; API changelog; GPT-6 Astra model page; GPT-6 Astra System Card; deprecations.
- Anthropic: models overview; release notes; model deprecations.
- Google: Gemini API models, changelog and deprecations; Gemini 3.8 Flash and Gemini 3.1 Pro model pages.
- xAI: API models and release notes; “xAI joins SpaceX” (2 February 2026), x.ai.
- Meta: “Introducing Muse Spark” (8 April 2026), about.fb.com; “Introducing Muse Spark 1.3” (2 September 2026), research.meta.ai; Meta Model API and Muse Glimmer pages, developer.meta.com.
- DeepSeek: API news, change log and pricing pages; DeepSeek-V4-Pro-0813 model card, Hugging Face.
- Alibaba Qwen: Qwen Cloud model changelog; Qwen3.8-2.4T-A95B model card, Hugging Face.
- Mistral AI: Mistral Medium 3.5 documentation; Mistral-Medium-3.5-128B model card, Hugging Face.
- Moonshot AI: Kimi-K3 model card, Hugging Face.
- Z.ai: release notes; GLM-5.3 guide.
Checked September 2026.
Related: What is a frontier AI model? · How large language models work
- artificial intelligence
- ai models
- explainer