Labs
The 41 AI labs the atlas follows, what each is betting on, how it is funded and what it has shipped.
Every funding round, newest first
Launched in the last 90 days
- Anthropic
Latest launch Claude Haiku 5.5, 7 October 2026
Frontier models as autonomous workers (coding, agents, science), released behind tiered safeguards (Fable and Mythos share one model with different access), and… - Google DeepMind
Latest launch SynthID Detector, 7 October 2026
AGI by pushing several fronts at once, across Gemini frontier models, agents, coding, robotics and world models, with science (AlphaFold, Isomorphic) as proof points.… - OpenAI
Latest launch GPT-6 with Intelligent UI in ChatGPT, 7 October 2026
Automate AI research. OpenAI's 'North Star' is a fully automated AI researcher (a research intern by 2026-09, declared met, and a multi-agent researcher by 2028-03),… - Meta Superintelligence Labs (MSL)
Latest launch Personal Agent Protocol, 6 October 2026
Rebuild the stack after Llama 4: closed frontier 'Muse' models with pretraining, RL and test-time scaling, a fast compute ramp, and consumer agents; open weights only… - Mistral AI
Latest launch Mistral Large 4, 6 October 2026
Sovereign, open-weights European frontier AI: train bigger sparse models, own data-center capacity (plan to roughly double compute ownership), and build custom… - Reflection AI
Latest launch Beam, 5 October 2026
An American 'open frontier' lab. The plan is to pretrain a large MoE open-weights model on tens of trillions of tokens, add reinforcement learning, and sell to… - Microsoft AI (MAI)
Latest launch MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and Voice-2.1-Flash, 1 October 2026
'Humanist superintelligence' means building Microsoft's own frontier MAI models for cost and independence from OpenAI, plus grounded, controllable systems for medicine… - DeepSeek
Latest launch DeepEP-Ascend and DeepGEMM-Ascend, 29 September 2026
Efficiency-first open frontier models, with sparse and compressed attention for 1M-token context at a fraction of compute, MoE at 1.6T parameters, aggressive pricing… - World Labs
Latest launch World Labs to join AMD, 28 September 2026
Spatial intelligence through generative and reconstructive world models that produce persistent, explorable 3D worlds and simulate space-time, aimed at gaming, VFX,… - Alibaba Qwen (Tongyi)
Latest launch Qwen3.8-Max recursive self-improvement demos (Apsara 2026), 22 September 2026
Largest open-weights ecosystem (460+ open models, 3B downloads) plus a proprietary Max flagship; scale to 5-10T parameters, push recursive self-improvement, and tie… - SpaceXAI (formerly xAI)
Latest launch Grok 4.7, 21 September 2026
Scale compute (Colossus, a reported million-H100-equivalent cluster; orbital data centers cited as the merger's main goal) and pair Grok with Cursor's coding product to… - Z.ai (Zhipu AI)
Latest launch GLM-5.3-Flash, 26 August 2026
Open-weights GLM frontier models and a domestic-compute story. GLM-5.3-Flash is claimed to serve all traffic on 100,000 Chinese-made chips. - MiniMax
Latest launch MiniMax Music 3.0, 13 August 2026
Commercial multimodal and coding models for consumers and developers (flagship M3), monetization-led growth. - ByteDance Seed
Latest launch SeedRealtime, 5 August 2026
Four 2026 priorities are world models (internal goal of a release by year-end benchmarked against Genie 3), keeping Seedance video near the top, a stronger coding… - Moonshot AI (Kimi)
Latest launch Kimi K3, 16 July 2026
Open-weights agentic frontier models trained cheaply, with Kimi Delta Attention (linear attention), the Muon optimizer, Agent Swarm parallel sub-agents, and 1M context…
Launched before that
- Periodic Labs
Latest launch The Art of Scaling RL Compute (ScaleRL), 15 October 2025
AI scientists plus autonomous physical labs. Robots run thousands of physics and chemistry experiments, and the experiment data (including negative results) trains… - Thinking Machines Lab
Latest launch Tinker, 1 October 2025
Customizable AI that people shape, through open-weights frontier models plus Tinker, a fine-tuning/RL training API, and 'interaction models' that process audio, video… - Inception (Inception Labs)
Latest launch Mercury (diffusion LLMs), 17 June 2025
Diffusion language models (the Mercury family) generate text by iterative refinement, with far lower latency and cost than token-by-token generation. Mercury 2…
No launch in the atlas yet
- Advanced Machine Intelligence (AMI Labs)
World models that learn abstract representations of real-world sensor data and predict in representation space (JEPA), not pixels or tokens, with persistent memory,… - Core Automation
New architectures and learning algorithms that need about 100x less data and learn continuously from experience, which founder Jerry Tworek sees as essential for AGI;… - Discovery Loop
Automate the scientific loop with AI systems that design and run thousands of experiments in parallel and improve themselves recursively, applied to hardware design,… - Elorian
Visual AGI: foundation models for visual understanding and reasoning, an area founder Andrew Dai (ex-Google DeepMind, Gemini) says has seen uneven progress. - Essential AI
Open models with better pretraining efficiency rather than heavy post-training; uses the Muon optimizer. Rnj-1 is an 8B open model aimed at coding and agents. - Eureka Labs
AI-native education, with AI teaching assistants alongside human-designed courses and Karpathy's educational material as the seed. - Flapping Airplanes
Data-efficient AI that learns more like humans do, which the company estimates are 100,000 to 1,000,000 times more data-efficient than current models; may require new… - General Intuition
World models and action-conditioned agents trained on a huge corpus of game footage from Medal (its parent video platform), generating synthetic training footage for… - humans&
AI built around how people collaborate. Its models and software act as connective tissue for teams and use reinforcement learning, memory and user understanding. It… - Ineffable Intelligence
Superintelligence from reinforcement learning on the system's own experience, without human text. The aim is a 'superlearner' that discovers knowledge from motor skills… - Isomorphic Labs
An AI drug design engine (IsoDDE) that goes beyond AlphaFold-style structure prediction to design molecules across therapeutic areas and modalities; Alphabet spinout… - Lila Sciences
'AI Science Factories' combine AI models with robotics, sensors and experimental feedback loops that form hypotheses, run tests and iterate. The aim is scientific… - Magic
Frontier code models that automate software engineering and AI research, the company's stated path to safe AGI: frontier-scale pretraining plus domain RL, ultra-long… - Mirendil
Sell the self-improving AI the big labs keep in-house, as an AI researcher platform that designs experiments, tunes settings, evaluates models and runs the next… - Physical Intelligence (Pi)
One general-purpose vision-language-action foundation model that controls many robot types, trained on diverse robot data plus web knowledge, improving through… - Poolside
Frontier coding and software-engineering models built from scratch, using a 'Model Factory' pipeline; a small team (fewer than 70 people built its model, per CEO).… - Pragmatik Labs
Next-generation agents across digital and physical worlds, with digital agents for knowledge work, physical agents for real environments, plus frontier research.… - Prometheus (Project Prometheus)
An 'artificial general engineer', meaning software that automates the design and manufacturing of complex physical systems, from jet engines to drug compounds. Bezos… - Recursive Superintelligence
AI that improves AI: a system that finds its own weaknesses and redesigns itself using open-ended algorithms, first automating AI research (ideation, implementation,… - Ricursive Intelligence
AI that designs chips and continues the AlphaChip line. It builds self-improving chip-design systems and positions itself as a neutral vendor chipmakers can trust with… - Safe Superintelligence Inc. (SSI)
A 'straight shot' to safe superintelligence with no products on the way. Sutskever says the field is leaving the age of scaling for an age of research; the core… - Sakana AI
Nature-inspired, efficient and sovereign AI for Japan. The approach is small data-efficient models, evolutionary model merging, and now orchestration of many frontier… - Unconventional AI
A new, energy-efficient computer for AI, aiming at biology-level efficiency (claimed 1,000x lower power in one report); hardware-software co-design instead of more GPUs.