The Labs
Building AI
Twelve frontier and open-source AI labs, organised by who builds the models. Each section: editorial intro, flagship models, what they bet on, what to use them for, and the catch.
OpenAI
The mainstream face of generative AI
Founded as a non-profit and now a commercial juggernaut, OpenAI turned ChatGPT into the consumer face of generative AI. The strategy: ship the broadest, most distribution-friendly model line on the planet.
Anthropic
The safety-first frontier lab
Spun out of OpenAI in 2021, Anthropic positions itself as the safety-conscious frontier lab. Claude is the model developers reach for when they want long context, careful reasoning, and an assistant that pushes back when it should.
Google DeepMind
Research giant turned product lab
The unified AI arm of Alphabet. Gemini powers Search AI Overviews, Workspace, Android, and the standalone Gemini app. Lean on the world's biggest distribution moat plus 25 years of research.
Meta AI
Open-weights at frontier scale
The Llama series became the open-weights frontier. Mark Zuckerberg's bet is that open models commoditise the layer above the chip, making Meta's app distribution and Reality Labs hardware more valuable.
xAI
The maximally curious lab
Elon Musk's frontier lab, with the implicit edge of X (formerly Twitter) as a real-time data feed. Four major Grok versions in under two years. Less hedged on political and controversial topics than competitors.
Mistral AI
European frontier lab
Paris-based, founded by Meta and DeepMind alumni. Hybrid strategy: open weights for some models, commercial APIs for the frontier. Strong on European languages, structured outputs, and shipping fast.
Qwen (Alibaba)
China's open frontier challenger
Alibaba's frontier AI lab, the most prolific releaser of open-weights models in 2025-2026. Strategy is volume: ship a new variant every six weeks, from 0.5B edge models to 200B+ flagships.
DeepSeek
The lab that proved compute isn't everything
Shocked Silicon Valley in early 2025 by training a frontier-class model for a tiny fraction of OpenAI's reported cost. Their mixture-of-experts architecture became the template the open-source world copies.
Cohere
Enterprise-first, retrieval-native
Founded by Aidan Gomez, one of the Transformer paper authors. Bet on enterprise from day one rather than consumer chat. The Command family is built for RAG, structured outputs, and on-premise deployment.
Perplexity
Search-native AI
Reframed AI as a search interface rather than a conversation. The product cites every claim back to a source. The Sonar model family is tuned for grounded, citation-aware answering.
Perceptron
Spatial intelligence for the physical world
Newer entrant focused on spatial intelligence — models that understand the physical world, not just text. Heavy weighting toward 3D scene understanding, robotics datasets, and embodied trajectories.
Nous Research
Open-weights craft house
An independent research collective rather than a venture-backed lab. Hermes Agent became one of the most-downloaded open agentic models of 2025-2026. The vibe is craftsmanship over scale.
