Thore Graepel, a co-creator of AlphaGo, launched Metis Reasoning in London on October 20, 2026: a 12-person lab that says AI reasoning should be checkable and far cheaper in compute than today's frontier models. The details come from Tech.eu's launch report and a follow-up summary from Resultsense. Some aggregator headlines go further and say the lab will "end AI hallucinations." The founders' own quoted words do not say that.
This post separates what Metis has claimed, what has been reported by outlets, and what has not been demonstrated. It also places the lab among the other new London research companies founded by senior researchers.
TL;DR: claimed vs verified
| Item | Status |
|---|---|
| Graepel left Google DeepMind this summer and founded Metis Reasoning | Reported by Tech.eu; launch dated October 20, 2026 |
| 12-person team of researchers, mathematicians, robotics experts, entrepreneurs | Reported by Tech.eu |
| Raising tens of millions of dollars | Reported by Bloomberg (September 24) and echoed by Tech.eu as "understood"; the company declined to confirm |
| Approach is more verifiable, robust and safer than current frontier models | Claimed by Metis; no evidence published |
| Architecture is "much much more compute-efficient" | Claimed by co-founder Piers Ingram; no figures |
| "Encouraging results" on frontier reasoning problems | Claimed; no benchmark named |
| Ends hallucinations | Not claimed by the founders in the quotes we saw; an aggregator framing |
| First public model date | Unknown per Tech.eu |
A maze illustrating step-by-step AI reasoning paths
What Metis says it is building
Tech.eu describes Metis as pursuing a "compute-efficient" alternative to the path Anthropic and OpenAI follow with frontier models. The pitch has three parts:
- LLMs reason unreliably. Metis grants that large language models now produce fluent answers, infer meaning, follow instructions and handle some multi-step tasks. Its argument is that they have a fundamental limitation in how they reason and leave no audit trail.
- That limitation blocks adoption. In engineering, science, medicine or law, the company says, use remains constrained by the difficulty of verifying the model's work. Ingram put it bluntly: in higher-stakes cases the arguments models produce "basically can't be trusted."
- AlphaGo's lesson applies. The thesis builds on Graepel's AlphaGo work, saying a similar architectural lesson should be applied to reasoning. Metis expects the approach to suit robotics and engineering.
On compute, Ingram said the team is "very much not going in this direction of ever larger models and ever larger training budgets" and that the architecture is "much much more compute-efficient" than frontier models.
Notice what is missing: no description of the architecture, no named benchmark, no number. The pitch is a thesis, not a result.
Where the "ends hallucinations" framing comes from
Some aggregator summaries of the launch say Graepel's lab will eliminate hallucinations. Neither Tech.eu nor Resultsense, the two reports we could read, quotes the founders saying that. Tech.eu says only that Metis argues its approach is "more verifiable, robust and safer" than current frontier models.
That is the company's aspiration, not a measured claim. Eliminating hallucinations is a much stronger statement than offering an audit trail. A system can show its steps and still be wrong, and a system can be right without showing steps. The honest reading of the evidence so far is "Metis wants reasoning whose steps can be inspected." Whether that reduces errors is an empirical question nobody can yet answer.
Who is behind it
According to Tech.eu:
- Thore Graepel: AlphaGo co-creator and former Google DeepMind research lead, who left the company this summer. A Bloomberg-based summary describes him as a machine-learning researcher who helped build AlphaGo.
- Piers Ingram: a serial technology and biotech entrepreneur who has raised more than 350 million dollars. He says he met Graepel at a Cambridge barbecue where they discussed Turing machines.
- Michael Warren: robotics and autonomous systems expert with Boeing experience.
- David Berman: head of AI at Cambridge Consultants, who has advised the UK government and NATO.
Tech.eu says the team spans researchers, mathematicians, robotics experts and tech entrepreneurs. The skew toward robotics and engineering experience matches the stated target markets.
Funding: what is reported and what is not
Bloomberg reported on September 24, 2026 that Graepel was seeking tens of millions of dollars for a reasoning startup, with the initial financing sought from a small group of backers, according to people familiar. The Runtime Wire summary adds that the amounts are targets, not closed funding. At launch, Tech.eu repeats that the company is understood to be raising tens of millions, and Ingram declined to comment on the reports, saying only: "We have been able to fund the company more than adequately to date."
So treat the funding figure as a reported target. Funding is not the story here, though. What matters to readers who build with AI is whether the technical bet is credible, and for that the public record is thin.
How this fits the wave of London research labs
Metis joins a cluster of new London labs started by senior researchers:
- Ineffable Intelligence, founded by Graepel's former AlphaGo colleague David Silver, pitches reinforcement learning as an alternative to LLMs. We covered its thesis in David Silver's Ineffable and the superlearner idea.
- Recursive Superintelligence, co-founded by Richard Socher, per Tech.eu.
Asked about overlap, Ingram said, "We obviously don't know exactly what they are doing; they don't know exactly what we are doing," and described Metis's approach as "really a very different, almost philosophical approach."
The common thread is a bet that scaling transformers is not the only path, and that search, planning and verification, the ingredients behind AlphaGo, belong in general-purpose systems. For context on where the large labs think capability is headed, see our coverage of DeepMind's paper on pathways from AGI to ASI and Geoffrey Hinton's intelligence explosion paper.
Why "verifiable reasoning" matters to builders
Whatever Metis ships, the problem it names is real, and anyone deploying AI at work feels it. Reviewing and redoing AI output has a measurable cost, which we looked at in the verification tax surveys. A system that returns a chain of checkable steps could cut that cost, if the steps are faithful and the checking is cheap.
Related research is already moving in this direction. Apple's work on selection-based structured reasoning for multimodal search agents is one example of constraining how models reason to make behavior more predictable.
What AlphaGo actually taught, and what may not transfer
AlphaGo combined learned judgment, a neural network that suggests promising moves and evaluates positions, with explicit search that looks ahead through possible continuations. The search step is why its play could be examined: each candidate line had a value the system computed. Go also has a perfect referee. The rules decide who won, so every training signal is unambiguous.
That is the part that may not transfer. In engineering, medicine or law, there is rarely a rulebook that scores an argument automatically. A reasoning system can only be verified as far as there is something to verify against: a formal proof checker, a simulator, a unit test, a body of regulation. Metis's targets, robotics and engineering, are domains where simulators and physical constraints supply some of that referee. Open-ended legal or medical reasoning supplies much less.
This is not a criticism of the idea. It is the reason the first domain Metis chooses to demonstrate will be informative. A result on PhD-level science problems, where answers can be checked against known solutions, would say something different from a result on messy real-world documents.
What to watch for
A reader trying to judge Metis in the coming months can look for four things:
- A public benchmark result on a named reasoning test, with the compute cost and the baseline model stated.
- An independent evaluation. None has been announced; it would be the first non-company evidence.
- Technical documentation. Any paper or write-up that describes the architecture and an evaluation plan, rather than a position.
- A release. Tech.eu says it is unclear when the first model will be public. Until there is something to run, claims stay claims.
How to evaluate any "verifiable AI" pitch
- Ask what is verified: the final answer, each step, or the process that produced it.
- Ask who does the verifying: a formal checker, a second model, or a human.
- Ask for error rates against a baseline on the same tasks.
- Compare compute cost per solved problem, not parameter counts.
- Be wary of "ends hallucinations" language; look for measured reductions instead.
Bottom line
Metis Reasoning is a credible team with a clear thesis and an AlphaGo pedigree, launching into a moment when investors are funding alternatives to scaling LLMs. It has not yet shown a model, a benchmark or a release date. Take the compute-efficiency and verifiability claims as hypotheses, and the "end of hallucinations" headline as an aggregator's gloss.
Related reading
- David Silver's Ineffable Intelligence and the superlearner idea
- DeepMind's AGI to ASI paper: four pathways
- Hinton's intelligence explosion paper
- The AI verification tax: surveys on rework and how to cut it
- Apple's selection-based structured reasoning for search agents
Accurate as of October 20, 2026. Claims are attributed to Tech.eu, Resultsense and Bloomberg-based reporting.