AI research labs

Organisations that train frontier models or publish foundational AI research, as of 2026-10-05. Facts below are summarised from the company notes (each carries its own sources and fetch dates; most were refreshed on 2026-10-05) and are mostly Wikipedia or press based, so check the linked note before quoting a figure. Model-level detail: ai-model-providers.

US frontier labs

• OpenAI (openai): founded December 2015; reportedly converted to the public benefit corporation OpenAI Group PBC on 2025-10-28, controlled by the OpenAI Foundation (reportedly about 26%; microsoft about 27%). Per the vault note, the current lineup is the GPT-6 generation (Astra 2026-09-03, 6.1 Sol 2026-09-29); see gpt-5-and-gpt-6-family.
• Anthropic (anthropic): public benefit corporation founded 2021; per the vault notes, Claude Fable 5.1 (2026-09-01), Opus 5.5 and Sonnet 5.5 are the current tiers, plus an invite-only Mythos variant (anthropic-claude, claude-fable-5). Press-reported confidential IPO filing 2026-06-01. Responsible Scaling Policy: ai-safety-and-governance.
• Google DeepMind (deepmind-research-lab, parent alphabet): founded 2010, merged with Google Brain in 2023-04, CEO Demis Hassabis; Gemini 3.x (gemini-3-x-family, Gemini 4 reportedly announced 2026-09-30 with restricted early access), open-weight gemma-4, and AlphaFold, AlphaProof and AlphaEvolve science work.
• Meta (meta): Meta Superintelligence Labs (from 2025-06, led by Alexandr Wang) reportedly replaced the Llama line with the closed Muse family from 2026-04-08; Muse Glimmer 30B open weights reportedly followed on 2026-08-10 (meta-muse).
• xAI / SpaceXAI (xai): founded 2023-03-09; reportedly acquired by SpaceX on 2026-02-02 and renamed SpaceXAI on 2026-07-06; Grok models (grok); SpaceX reportedly bought Cursor’s parent Anysphere for USD 60B (closed 2026-08-14, cursor).
• Thinking Machines Lab (thinking-machines-lab): founded February 2025 by Mira Murati; Tinker fine-tuning API (tinker) and open-weight Inkling models from July 2026 (per the company note; press-based).

Europe and China

• Mistral AI (mistral-ai): Paris, founded April 2023; reported EUR 3B Series D on 2026-09-08 at over EUR 21B; open-weight Mistral Large 3 and Small 4, Vibe assistant (mistral-family).
• DeepSeek (deepseek): Hangzhou, founded 2023-07-17; MIT-licensed V4 open-weight models (deepseek-family). Its company note is still DRAFT and single-sourced.
• Alibaba / Qwen (qwen, alibaba, qwen-family): open and proprietary models; Qwen 3.8 generation; Qwen 4 reportedly previewed 2026-09-22 but not released.
• Z.ai (z-ai): GLM family, reportedly listed in Hong Kong since 2026-01-08; GLM-5.2 reportedly released 2026-06-16.
• Others: moonshot-dot-ai (kimi-k3), minimax, openbmb.

Open-source, tooling and enterprise labs

nous-research (Hermes), hugging-face (model hub; the company note records a reported Nvidia acquisition announcement on 2026-09-03), ibm (granite), Nvidia (nvidia-nemotron), apple.

Context

Stanford’s 2026 AI Index reports industry produced over 90% of notable frontier models in 2025 and that US and Chinese leaders have traded places since early 2025 (see future-trends-in-ai). Lab safety policies: ai-safety-and-governance. Funding and valuation figures change quickly and are carried only in the company notes.

Open items

• Primary-source confirmation (SEC filings, company announcements) for the SpaceX/xAI rename, Mistral’s September 2026 round, Meta’s Muse Glimmer and the Hugging Face acquisition: the company notes rely on Wikipedia and press.
• Not covered: Microsoft AI, Amazon, Apple’s own models, SSI, Reflection, AI2, Cohere, Zhipu competitors beyond those listed.

Sources

• The linked company notes in RESOURCES/COMPANIES/ (fetched 2026-10-05 by their own jobs, see each Sources section)
• https://hai.stanford.edu/ai-index/2026-ai-index-report (accessed 2026-10-05)