Frontier AGI labs
The fifteen industrial and non-profit laboratories whose model releases set the global state of the art.
A frontier lab is one that trains and publicly releases models at — or close to — the global capability frontier. The list below covers the labs that, as of 2026, have either trained a frontier model in the last 24 months or run compute clusters comparable in scale to those that have. Each entry links to the lab's research home page.
- 01
OpenAI
2015 · San Francisco, USACreator of the GPT family and o-series reasoning models. Operates a hybrid capped-profit structure with a stated mission to ensure broadly beneficial AGI.
Frontier LLMsReasoningAgents - 02
Google DeepMind
2010 · London, UKUnified Google's AI research after the 2023 merger with Google Brain. Builds the Gemini family and breakthrough scientific systems such as AlphaFold and AlphaProof.
RLScienceMultimodal - 03
Anthropic
2021 · San Francisco, USASafety-focused lab founded by former OpenAI researchers. Develops the Claude model family and publishes widely on Constitutional AI, alignment, and mechanistic interpretability.
SafetyInterpretabilityClaude - 04
Meta FAIR
2013 · Menlo Park, USAMeta's Fundamental AI Research team. Releases the Llama family of open-weight models and pursues long-horizon research on world models, robotics, and self-supervised learning.
Open weightsMultimodalEmbodied AI - 05
xAI
2023 · Palo Alto, USAElon Musk-founded lab building the Grok model family. Trains on the Memphis 'Colossus' GPU cluster, one of the largest single-site training installations in the world.
LLMsReasoningReal-time data - 06
Mistral AI
2023 · Paris, FranceEuropean frontier lab specialising in efficient open-weight and commercial models. A focal point for sovereign AI capability in the EU.
Open weightsEfficient models - 07
Cohere
2019 · Toronto, CanadaEnterprise-focused lab building retrieval-augmented LLMs and the Command family. Co-founded by Aidan Gomez, one of the authors of the original Transformer paper.
Enterprise LLMsRAG - 08
Microsoft Research AI
1991 · Redmond, USAOne of the oldest industrial research labs. Drives Microsoft's Phi family of small models and collaborates closely with OpenAI on frontier deployment.
FoundationsAI for science - 09
Apple Machine Learning Research
2017 · Cupertino, USAPublishes Apple's on-device and Apple Intelligence research, with a strong emphasis on efficient, privacy-preserving models that run on consumer hardware.
On-device MLPrivacy - 10
Allen Institute for AI (AI2)
2014 · Seattle, USANon-profit research institute founded by Paul Allen. Maintains the fully open OLMo model family and the Semantic Scholar academic search engine.
Open scienceReasoning - 11
Mila
1993 · Montréal, CanadaQuebec AI Institute, founded by Yoshua Bengio. One of the world's largest academic deep-learning groups.
Deep learningAI for good - 12
Vector Institute
2017 · Toronto, CanadaIndependent research institute anchoring Toronto's AI ecosystem, with strong ties to Geoffrey Hinton and the University of Toronto.
Applied MLHealth AI - 13
Tsinghua KEG / IIIS
1996 · Beijing, ChinaTsinghua University's flagship AI groups. Released the GLM and ChatGLM model families and trains many of China's leading AI researchers.
LLMsKnowledge graphs - 14
BAAI
2018 · Beijing, ChinaBeijing Academy of Artificial Intelligence. State-affiliated lab coordinating much of China's foundation-model and AI-governance research.
Foundation modelsAI policy - 15
RIKEN AIP
2016 · Tokyo, JapanRIKEN's Center for Advanced Intelligence Project. Japan's central public AI research lab, with a strong theoretical and scientific bent.
ML theoryAI for science
How to use this list: skim once to get a map of who trains what; bookmark the research pages of the labs whose work most affects your industry; revisit quarterly when major model releases land.