Summary: Google DeepMind’s February 11, 2026 update positions Gemini 3 Deep Think for research-grade mathematics, physics, and computer science rather than only olympiad problems. The Aletheia math agent can admit failure, verifies natural-language proofs, and can draw on web search; Deep Think reaches up to 90% on IMO-ProofBench Advanced as inference compute increases. Results span 18 research problems with at least one ICLR 2026 acceptance, while access remains limited to AI Ultra and an Early Access Program.
February 11, 2026 was the day Google DeepMind chose to publish a substantial Gemini 3 Deep Think update on its official blog. This specialized reasoning mode has moved past the goal of solving olympiad problems faster; it now exists to take on open-ended research questions in mathematics, physics, and computer science under the direction of professional researchers. Thang Luong and Vahab Mirrokni led the post, Demis Hassabis and Jeff Dean sponsored it, and Terence Tao appears among the thanked external experts. A byline list like that carries its own message: at DeepMind, “AI for Science” is a frontline priority, not a side project.
Past Olympiad Gold, Into Research-Grade Reasoning
Within Gemini 3, Deep Think is the dedicated reasoning mode. By DeepMind’s account, it hit gold-medal standard at the International Mathematical Olympiad in summer 2025, and a later version matched that level at the International Collegiate Programming Contest. The January 2026 release is what this update centers on: internal evaluations have it clearly beating the IMO-gold version while spending less inference-time compute.
Capability, in other words, keeps climbing while the cost curve bends downward — and for products built on reasoning modes, that trend matters more than any single leaderboard score.
Aletheia: The Math Agent That Can Admit It Failed
For engineers, the most interesting piece here is Aletheia, a mathematical research agent with Deep Think underneath. The agent runs a natural-language verifier over its own proofs, leans on Google Search and web browsing for literature work, and carries an explicit design decision: it is allowed to admit failure.
That last property matters enormously in agent engineering. A research agent with no way to say “I could not do this” is more dangerous than one that fluently produces wrong proofs. And because verification happens in natural language, mathematicians get to audit machine reasoning in the format they already work in instead of being forced through formal proof systems.
Reasoning That Scales: Up to 90% on IMO-ProofBench Advanced
Raise the inference-time compute and Deep Think’s IMO-ProofBench Advanced score climbs with it, up to 90%; DeepMind also reports that the scaling law holds through PhD-level exercises on its internal FutureMath Basic benchmark. Read as a product statement, this is direct: once quality can be bought with compute, accuracy, latency, and cost collapse into one set of tunable dials, and pricing can be segmented by scenario rather than shipped as a single fixed model property.
Eighteen Research Problems, One Concrete Ledger
Concrete ledgers are rare in this field, but DeepMind published one: across 18 research problems, the work produced at least one ICLR 2026 acceptance. Highlights include Feng26, a fully autonomous paper computing eigenweight structure constants in arithmetic geometry; autonomous solutions to 4 of the 700 open problems in Bloom’s Erdős conjecture database (including Erdős-1051, later generalized in BKKKZ26); and a three-item counterexample that refutes a 2015 conjecture in online submodular optimization. The team also deployed it to assist STOC 2026 paper reviews.
Just as notable is the underclaiming. The report states explicitly that results so far are Level 2 “publishable quality,” with no Level 3 or Level 4 breakthroughs — and in a hype-driven release culture, that restraint makes the whole report more credible, not less.
Where It’s Available, and Where It Isn’t
Deep Think currently lives in the Gemini app, open to Google AI Ultra subscribers — Google confirmed as much through its official channels. The API side is narrower: selected researchers, engineers, and enterprises can apply through the Early Access Program, and the general pay-as-you-go Gemini API remains out of reach. Demis Hassabis followed up on X to announce that the upgrade set new records on “the most rigorous benchmarks in maths and science,” while India Today’s later coverage reported passing marks on Humanity’s Last Exam. Teams that need it now have one reliable path — the subscription tier; the API route means applying and waiting.
Three Takeaways for Developers
First, verifier design: pairing natural-language verification with permission to fail gives you a reusable agent architecture pattern for any high-stakes generation pipeline. Second, inference compute as a product variable: that 90% figure sits on a controllable quality-cost curve, one that belongs in your own eval and pricing design. Third, the differentiation battlefield is shifting from generic benchmarks to professional research workflows. Picking up the acceleration and roadmap-divergence threads from our 2026 opening outlook, Deep Think bets on going deeper while the open-weight camp pushes prices down — and the split between those two routes will only get sharper.