There is no agreed test showing that any company has built artificial general intelligence, so a list of "AGI companies" really means firms that state it or something near it as a goal and publish documents about how they will manage the risks. Three of those documents can be checked directly: Google DeepMind's Frontier Safety Framework (version 3 published September 22, 2025), the Frontier AI Safety Commitments that 16 companies signed on May 21, 2024, and the self-descriptions on company websites.
Key takeaways
- Companies describe their goals in different words: general intelligence, advanced AI, or safe superintelligence.
- The Seoul commitments of May 21, 2024, were voluntary, and 20 organizations had signed by the time the UK government updated its page on February 7, 2025.
- Google DeepMind's framework defines Critical Capability Levels, thresholds at which a model may pose heightened risk of severe harm without mitigations.
- A company's published framework shows what it plans to do, not what an outside party has verified.
- This article ranks no company and takes no position on who is closest to AGI.
How three organizations describe their aims
The table paraphrases pages that each organization publishes about itself. Self-descriptions are statements of intent, and they change, so the date checked matters.
| Organization | What its own page says | Structure stated on the page |
|---|---|---|
| Google DeepMind | Mission is to "build AI responsibly to benefit humanity"; says AI, and ultimately AGI, could drive major change | Not stated in the page text reviewed |
| Anthropic | Mission concerns a safe transition through advanced AI; the term AGI does not appear on the company page | Public benefit corporation, with a board elected by stockholders and overseen by a Long-Term Benefit Trust |
| Safe Superintelligence Inc. | Describes building a safe superintelligence as its sole aim | American company with offices in Palo Alto and Tel Aviv |
The wording matters for readers comparing headlines. A company that targets "superintelligence" is claiming something different from one that says "general" capability, and a company that avoids both terms is making a narrower public claim. Other firms, including OpenAI, are left out of the table because their own pages could not be retrieved for this check.
The Seoul commitments
On May 21, 2024, the UK government announced that 16 companies had agreed to the Frontier AI Safety Commitments at the AI Seoul Summit. The published text says signatories commit to publish a safety framework focused on severe risks, to set thresholds at which risks would be intolerable unless mitigated, and to explain how those thresholds were chosen. In the most extreme cases they commit not to develop or deploy a model at all if risks cannot be kept within thresholds.
The page, last updated February 7, 2025, lists 20 signatories: Amazon, Anthropic, Cohere, Google, G42, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, the Technology Innovation Institute, xAI, Magic, Minimax, and NVIDIA. The commitments are voluntary, and the text does not describe an outside body that checks them.
What a published threshold looks like
Google DeepMind's blog post of September 22, 2025, describes the third version of its Frontier Safety Framework. Critical Capability Levels are capability thresholds at which, without mitigations, frontier models may pose heightened risk of severe harm. Version 3 added a level for harmful manipulation, protocols for models that could accelerate AI research and development to potentially destabilizing levels, and coverage of cases where a misaligned model might interfere with operators' ability to direct, modify, or shut it down.
When a level is reached, the company says it conducts safety case reviews before external launches, and that for the research-acceleration level these reviews also cover large-scale internal deployments. The post was later updated on April 17, 2026, when version 3.1 added Tracked Capability Levels. The framework is a company document, so it shows a commitment on paper rather than an audit result.
Questions for reading an AGI announcement
- Does the company define the term it uses, and where is that definition published?
- Is there a public framework with named thresholds, and which version is current?
- Did an outside evaluator test the claim, or is it the company's own measurement?
- Is the statement about a goal, a research result, or a product on sale?
- Does the corporate structure give anyone other than shareholders a role in oversight?
FAQ
Which company has built AGI?
None has been shown to have done so in the sources checked as of October 2026. The term has no single accepted definition, so any such claim depends on which definition is used.
Are the Seoul commitments legally binding?
The government text presents them as voluntary commitments by companies. It does not describe penalties or an enforcement body.
What is a Critical Capability Level?
In Google DeepMind's framework, it is a capability threshold at which a frontier model may pose heightened risk of severe harm unless mitigations are applied.
Is superintelligence the same as AGI?
The terms are used differently by different organizations, and no standard definition separates them across the sources checked. A company that names superintelligence as its aim is stating a different goal than one that names general capability.
Related reading
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Frontier AI Safety Commitments, AI Seoul Summit 2024, UK Government, May 21, 2024 (page updated February 7, 2025)
- Frontier Safety Framework update post, Google DeepMind, September 22, 2025 (updated April 17, 2026)
- About Google DeepMind, Google DeepMind, 2026
- Company page, Anthropic, 2026
- Safe Superintelligence Inc. home page, 2026