The Labs: Who's Who
Table of Contents
Link copied!
Snapshot as of July 19, 2026. Valuations and funding are dated and sourced. 🔄
How this is organized: grouped by what developers actually reach for each lab for. The Open/closed column tells you whether you can download the weights and run them yourself (open) or only reach the model through an API (closed). Several labs do both, so it reflects their main stance.
| Lab | HQ | Open/closed | Flagship (Jul 2026) | Website | Backing / latest valuation |
|---|---|---|---|---|---|
| OpenAI | USA | Closed | GPT-5.6 family | openai.com | ~$852B (Mar 2026) |
| Anthropic | USA | Closed | Claude Fable 5 (+ Opus 4.8 / Sonnet 5) | anthropic.com | ~$965B (May 2026) |
| Google DeepMind | UK / USA | Mixed | Gemini 3.1 Pro (+ 3.6 Flash, 3.5 Flash-Lite, Gemma) | deepmind.google | Alphabet subsidiary |
| xAI | USA | Closed | Grok 4.5 | x.ai | Part of SpaceX (~$1.25T combined) |
| Microsoft AI | USA | Mixed | MAI-Thinking-1 (+ Phi) | microsoft.com/ai | Microsoft division |
| Amazon | USA | Closed | Nova 2 | aws.amazon.com/nova | Amazon division |
| Meta (MSL) | USA | Open → closed | Muse Spark 1.1 (closed) + open Llama 4 | ai.meta.com | Meta division |
| Mistral | France | Open + closed | Medium 3.5 + Large 3 (both open) | mistral.ai | ~€11.7B (2025) |
| DeepSeek | China | Open (MIT) | DeepSeek V4 | deepseek.com | >$50B (Jun 2026) |
| Alibaba (Qwen) | China | Open + closed | Qwen3.8-Max (open line: Qwen3.6) | qwen.ai | Alibaba division |
| Moonshot (Kimi) | China | Open | Kimi K3 (2.8T) | moonshot.ai | ~$20B (May 2026) |
| Zhipu / Z.ai | China | Open | GLM-5.2 | z.ai | Public, HK IPO ~$6.6B (Jan 2026) |
| Cohere | Canada | Mixed | Command A+ / Embed 4 | cohere.com | ~$6.8B (2025) |
| AI21 Labs | Israel | Hybrid | Jamba | ai21.com | ~$1.4B (2023) |
| Reka | USA | Hybrid | Reka Flash / Core | reka.ai | ~$1B (2025) |
Frontier labs
- OpenAI. The default general-purpose API and the largest tooling ecosystem.
- Anthropic. A leading model for coding and agentic/tool-use workloads; safety-focused.
- Google DeepMind. The largest context windows and deep integration across Cloud, Android, and Workspace.
- xAI. Real-time, X-connected models; now part of publicly-traded SpaceX (branded "SpaceXAI") after the 2026 merger.
- Microsoft AI. First-party "MAI" models on Azure/Foundry, alongside its OpenAI partnership. It also ships the open Phi small models.
- Amazon (Nova). Price-performance models tightly integrated with AWS and Bedrock.
Open-weight leaders
- Meta (MSL). Llama is still a self-hosting/fine-tuning default, but the lab is pivoting toward closed frontier models (Muse Spark 1.1).
- Mistral AI. The go-to European, permissively-licensed, self-hostable option.
- DeepSeek. Frontier-adjacent quality at rock-bottom cost under a permissive MIT license.
- Alibaba (Qwen). The most common open base for fine-tuning and edge deployment, with the broadest range of sizes.
- Moonshot (Kimi). Frontier-competitive open weights for coding and agents (Kimi K3).
- Zhipu / Z.ai (GLM). Strong open coding and agentic models from a now-public, well-capitalized vendor.
Enterprise & specialist
- Cohere. Strong retrieval/embedding models and on-prem deployment for regulated industries.
- AI21 Labs. Efficient long-context (Jamba) for enterprise RAG.
- Reka. Deployable, efficient multimodal models plus an openly downloadable reasoning model.