AI reasoning
2 labs · 7 models and products Theme: Companies · Generated from this page's data. Click a name to open its panel on the Labs page. Records resting on one source are tagged "one source".AI reasoning means getting AI systems to work through problems step by step, as in mathematical proofs, logic or planning, and to reach answers that can be checked. Google DeepMind's AlphaGeometry pairs a language model with a logic engine to prove Olympiad geometry problems (per the site), and an advanced version of Gemini Deep Think reached gold-medal standard at the 2025 International Mathematical Olympiad. Metis Reasoning, a new company, says its AI is "designed to produce conclusions and take actions that are grounded, understood, and trusted", and aims to need far less computing power (per the site).
Context
- AlphaGeometry, a Google DeepMind theorem prover for Olympiad geometry, was published in Nature. 17 January 2024 (online; issue 18 January 2024) nature.com/articles/s41586-023-06747-5.pdf
We propose AlphaGeometry, a theorem prover for Euclidean plane geometry
- Google DeepMind reports that an advanced version of Gemini Deep Think reached gold-medal standard at the International Mathematical Olympiad in summer 2025. 11 February 2026 deepmind.google/blog/accelerating-mathematical-a
an advanced version of Gemini Deep Think achieved Gold-medal standard at the International Mathematics Olympiad (IMO)
- In an evaluation of 700 open Erdős problems, Google DeepMind reports Gemini Deep Think produced its own solutions to four of them. 11 February 2026 deepmind.google/blog/accelerating-mathematical-a
autonomous solutions to four open questions listed there.
- The AI Security Institute reports that in 2025 it tested the first model able to complete cyber tasks meant for experts with over ten years' experience. 18 December 2025 aisi.gov.uk/blog/5-key-findings-from-our-first-f
In 2025, we tested the first model that could complete cyber tasks intended for experts with over ten years of…
- Metis Reasoning describes its aim as AI whose conclusions and actions can be understood and trusted. accessed 31 October 2026 metisreasoning.ai/
Metis is developing AI designed to produce conclusions and take actions that are grounded, understood, and trusted.
Key papers
- Solving olympiad geometry without human demonstrations Nature · 17 January 2024
- Towards Autonomous Mathematics Research arXiv preprint · 10 February 2026
Labs Round 8 (31 Oct 2026): introduction and facts as the researchers wrote them; "(per the site)" marks a sentence resting on this site's own description of that lab.
| Record | What it is | Place | Links | Confidence |
|---|---|---|---|---|
| AlphaCode 2Chips, models and products | Second AlphaCode system for competitive programming, solving more contest problems than the first. research (technical report)Key facts Solved 1.7x more problems than AlphaCode; performed better than 85% of competition participants; technical report not opened (storage.googleapis.com blocked); access: paper onlyMade by Google DeepMind · 2023 (reported in DeepMind review dated 22 Dec 2023)Versions AlphaCode → AlphaCode 2 | London | deepmind.google/blog/2023-a-year-of-groundbreaki | one source |
| AlphaGeometryChips, models and products | A system that pairs a language model with a logic engine to prove Olympiad geometry problems. released (paper; open code and model)Key facts Theme: AI reasoning. Nature paper 17 Jan 2024; code and model open-sourced on GitHub. Figures (17 Jan 2024): solved 25 of 30 Olympiad geometry problems in time limits (previous best 10; average human gold medallist 25.9).Made by Google DeepMind · 2024-01-17Versions AlphaGeometry → AlphaGeometry 2Read the paper | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/alphageometry-an-olympiad-l · deepmind.google/blog/page/9/ | one source |
| AlphaGeometry 2Chips, models and products | Improved version of AlphaGeometry, pairing a Gemini-based language model with a symbolic engine to solve olympiad geometry. research (competition result)Key facts Announced 25 Jul 2024; solved 83% of IMO geometry problems from the past 25 years (predecessor 53%); proved IMO 2024 Problem 4 within 19 seconds of receiving its formalisation; access: paper onlyMade by Google DeepMind · 2024-07-25Versions AlphaGeometry → AlphaGeometry 2 | London | deepmind.google/blog/ai-solves-imo-problems-at-s | one source |
| AlphaProofChips, models and products | Reinforcement-learning system that proves maths statements in the formal language Lean; solved three IMO 2024 problems. research (paper; competition result)Key facts Announced 25 Jul 2024; with AlphaGeometry 2 solved 4 of 6 IMO 2024 problems, 28 of 42 points (silver-medal standard); AlphaProof solved two algebra and one number theory problem; methodology published in Nature 12 Nov 2025 (s41586-025-09833-y); access: paper onlyMade by Google DeepMind · 2024-07-25 | London | deepmind.google/blog/ai-solves-imo-problems-at-s | one source |
| DeepNash (Stratego agent)Chips, models and products | DeepMind agent that taught itself the hidden-information board game Stratego to human expert level. research (Science paper)Key facts Announced 1 Dec 2022; all-time top-three ranking among human experts on the Gravon Stratego platform; win rate above 97% (often 100%) against top Stratego bots; published in Science; access: paper onlyMade by Google DeepMind · 2022-12-01 | London | deepmind.google/blog/mastering-stratego-the-clas | one source |
| Gemini 2.0 Flash Thinking (experimental)Chips, models and products | First Gemini model released in a "thinking" mode that reasons step by step before answering. superseded (public preview)Key facts Public preview 19 Dec 2024; updated preview gemini-2.0-flash-thinking-exp-01-21 released 21 Jan 2025; access: API (preview); later Gemini 2.5 thinking models are listed as BIZ-LAB-7130Made by Google DeepMind · 2024-12-19 | London | ai.google.dev/gemini-api/docs/changelog | one source |
| Gemini 3 Deep ThinkChips, models and products | Gemini's most specialised reasoning mode, updated for science, research and engineering problems. released (updated reasoning mode)Key facts Theme: AI reasoning. Updated 12 Feb 2026; companion post on mathematical and scientific discovery with Deep Think 9 Feb 2026.Made by Google DeepMind · 2026-02-12 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/rss.xml · deepmind.google/blog/rss.xml | lab's own pages |
| Metis ReasoningCompanies | New company developing an AI reasoning system meant to be easier to check and to need far less computing power than today's large models.Profile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source (founders and headcount from trade only; the company site names no people). Sources Selected papers (1)
Key facts and news (7)
News
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | London | Companies House · tech.eu/2026/10/20/former-google-deepmind-top-re · metisreasoning.ai/ · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. | verified |
| Unlikely AICompanies · also on Software and tools | Builds its own AI for decisions in regulated fields such as financial services, accounting and insurance, where answers must be explainable and correct.Profile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source (all from company pages; seed round already verified in the deal file). Sources Models and products (1)
Key facts and news (6)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | London | Companies House · unlikely.ai/ · uktech.news/ai/unlikely-ai-funding-20220915 · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. | verified |