Labs
19 labs with a current model worth tracking, ordered by the strength of their best one.
Anthropic
San Francisco, United States
SI safety and research company, organised as a public benefit corporation, that builds the Claude models. Its lineup runs Fable, Opus, Sonnet and Haiku from most to least capable.
- Strongest model
- Claude Opus 5.5
- Intelligence Index
- 57.6
OpenAI
San Francisco, United States
Research and deployment company behind the GPT models, ChatGPT and Codex. Since GPT-5.6 its tiers carry names — Astra, Sol, Terra and Luna — while the number marks the generation.
- Strongest model
- GPT-6 Astra
- Intelligence Index
- 52.7
Google DeepMind
London, United Kingdom
Google's SI lab, formed by combining DeepMind with Google's Brain team. It builds the Gemini models and the open-weights Gemma family.
- Strongest model
- Gemini 4 Argon
- Intelligence Index
- 52.6
Meta
Menlo Park, United States
Meta's Superintelligence Labs group develops the Muse family: Muse Spark as the hosted line and Muse Glimmer as open weights.
- Strongest model
- Muse Spark 1.3
- Intelligence Index
- 48.1
SpaceXAI
Palo Alto, United States
Maker of the Grok models. Founded as xAI in 2023, it was acquired by SpaceX in February 2026 and now operates under the SpaceXAI name.
- Strongest model
- Grok 4.7
- Intelligence Index
- 46.4
Xiaomi
Beijing, China
Consumer-electronics company whose MiMo team releases large open-weight multimodal models under the MIT licence.
- Strongest model
- MiMo-V2.6-Pro
- Intelligence Index
- 46.3
Alibaba
Hangzhou, China
Alibaba's Qwen team develops the Qwen models, sold through Alibaba Cloud and partly released as open weights.
- Strongest model
- Qwen3.8-Max
- Intelligence Index
- 45.4
Z.ai
Beijing, China
Beijing lab, also known as Zhipu AI, that builds the GLM models and releases open weights for most of them.
- Strongest model
- GLM-5.3
- Intelligence Index
- 44.8
StepFun
Shanghai, China
Shanghai lab building the Step family of multimodal models.
- Strongest model
- Step 5 Preview
- Intelligence Index
- 43.7
Moonshot AI
Beijing, China
Beijing developer of the Kimi assistant and the Kimi K-series models, which it releases as open weights and serves through its own API.
- Strongest model
- Kimi K3
- Intelligence Index
- 43.6
DeepSeek
Hangzhou, China
Hangzhou lab that publishes its large models as MIT-licensed open weights and sells them through a low-priced first-party API.
- Strongest model
- DeepSeek V4.1 Flash
- Intelligence Index
- 39.5
Mistral AI
Paris, France
European lab that ships both open-weight and commercial models, and runs the Le Chat assistant and the Vibe coding agent.
- Strongest model
- Mistral Large 4
- Intelligence Index
- 38.4
MiniMax
Shanghai, China
Shanghai company that builds the M-series language models alongside video, speech and music models, and publishes weights for its main text models.
- Strongest model
- MiniMax-M3
- Intelligence Index
- 29.2
Thinking Machines Lab
San Francisco, United States
Research and product company behind the open-weight Inkling models and the Tinker fine-tuning platform.
- Strongest model
- Inkling-Small
- Intelligence Index
- 25.7
Tencent
Shenzhen, China
Tencent's Hy (Hunyuan) team builds open-weight language models, sold commercially through Tencent Cloud.
- Strongest model
- Hy3
- Intelligence Index
- 25.3
NVIDIA
Santa Clara, United States
Publishes the Nemotron family of open-weight models, along with training data and recipes, for building and customising agents on NVIDIA hardware.
- Strongest model
- Nemotron 3 Ultra
- Intelligence Index
- 22.9
Cohere
Toronto, Canada
Enterprise-focused lab behind the Command and North model families, with an emphasis on retrieval, tool use and private deployment.
- Strongest model
- Command A+
- Intelligence Index
- 13.1
Amazon
Seattle, United States
Amazon's in-house Nova family of text, multimodal and speech models, sold through Amazon Bedrock.
- Strongest model
- Nova 2 Lite
- Tracked
- 1 model
Microsoft
Redmond, United States
Microsoft AI trains the in-house MAI models for Copilot and Microsoft Foundry; Microsoft also publishes the open-weights Phi small-model family.
- Strongest model
- MAI-Thinking-1
- Tracked
- 1 model