Gemini Deep Think Moves From Benchmarks to Real Discovery
Gemini Deep Think has crossed from benchmark scores into real research, with Google DeepMind's Aletheia agent autonomously resolving four previously open Erdős problems out of 700 and contributing to published papers. A year after its 2025 IMO gold-medal run, the model is now producing publishable-quality mathematics and physics results with a human in the loop. The bigger signal: frontier reasoning is moving from passing exams to doing net-new expert work. For PMs, the defensible layer is shifting from the base model everyone rents to the verifier and problem-framing scaffold you build around it.
Ask yourself an uncomfortable question this week: if a frontier reasoning model can already produce publishable-quality work in your domain with a human checking it, what part of your product is the model doing, and what part is the scaffolding you built around it? That distinction is now your entire defensibility. Here is the concrete move: pull your three highest-value reasoning workflows and, for each, separate the generation step from the verification step on paper. Wherever you have generation but no independent verifier, you have both a quality ceiling and a copyability problem, because a better base model closes the first gap and a competitor's verifier closes the second. Build or at least specify that verifier this quarter. To be fair to the teams shipping thin wrappers, the ground moved fast, and eighteen months ago a dedicated checking layer looked like over-engineering rather than a moat. It isn't over-engineering now. Google's own results came from the natural-language verifier, not a bigger model, and they published the pattern for anyone to copy. The window to own the verification and problem-framing layer in your niche is open, but it closes the moment your category's obvious verifier ships as a default feature in the next model release. Move while the scaffolding is still the scarce thing.
Google DeepMind's Aletheia, built on Gemini Deep Think, autonomously resolved four previously open Erdős problems out of 700 attempted and produced 63 technically correct solutions overall.
Aletheia was pointed at 700 open problems from Bloom's Erdős conjecture database, the kind of questions professional mathematicians leave unsolved for decades, and it resolved four of them outright. Not "matched a known answer." Solved. Previously open.
That is the line worth staring at. For three years the story of frontier models in math was a story about tests: can the model pass the exam a strong high-schooler passes. As of mid-2026, according to Google DeepMind, the story has quietly changed.
The exam is over. These systems are now contributing to the literature.
My verdict: this is the moment reasoning stopped being a benchmark sport and became a research instrument. Slower than the hype implied, realer than the skeptics allowed.
From gold medals to open problems
Rewind twelve months. In July 2025, an advanced Gemini with Deep Think scored 35 out of 42 at the International Mathematical Olympiad, gold-medal standard and the first AI to earn it, per Google DeepMind. In September it matched that level at the International Collegiate Programming Contest.
Impressive, and ultimately still a test. The IMO has an answer key.
What DeepMind detailed this year is different in kind. Aletheia is a research agent built on Deep Think with a natural-language verifier that hunts for flaws in its own candidate proofs. It hit 90% on IMO-ProofBench Advanced, and more tellingly produced 63 technically correct solutions across those 700 Erdős problems, four of them genuinely new.
On Erdős-1051 it went further, helping generalize the result into a published paper. On the harder FutureMath benchmark of PhD-level exercises it managed 38%, a useful reminder of how much headroom remains.
The pattern repeats outside pure math. DeepMind reports Deep Think broke a decade-old conjecture in online submodular optimization by constructing a three-item counterexample, and cracked deadlocked Max-Cut and Steiner Tree problems by importing tools from continuous mathematics, the Kirszbraun theorem and measure theory, into discrete settings where humans hadn't thought to look. It extended auction theory's Revelation Principle to continuous values using topology, and settled cosmic-string physics integrals with Gegenbauer polynomials.
The transition nobody quite named
Here is the arc that matters. Every general-purpose technology moves through the same three phases: it beats the human at a game, then assists the human at work, then reshapes the work itself. Chess engines did it.
Compilers did it. The internet did it to distribution.
Reasoning models are mid-phase-two. DeepMind offers its own taxonomy, levels 0 through 4 from fully autonomous to collaborative, and places today's frontier at Level 2: human-guided but producing output of publishable quality. That is a precise, and notably modest, self-assessment.
It helps to picture what the system actually is. Less a machine mathematician than a tireless postdoc who has read every paper in every field, and who occasionally notices that a trick from topology dissolves the problem stuck in combinatorics. That cross-field recall, not raw autonomy, is where the four Erdős solutions came from.
To be fair to the skeptics, "publishable quality" is doing real work in that sentence, and four solved problems out of 700 is a low hit rate that leaned on human framing at every step. This is not an independent researcher. It is a very fast, very well-read collaborator.
But collaboration is exactly the capability that compounds, and the people driving it, Thang Luong, Vahab Mirrokni, Tony Feng and David Woodruff among them, are careful to claim assistance rather than replacement.
What it signals for anyone building on reasoning
If you ship a product that touches expert reasoning, legal analysis, drug discovery, quant research, code verification, chip design, the relevant question just changed. It is no longer "can the model pass a benchmark." It is "can the model do net-new work a domain expert would sign their name to." As of mid-2026 the honest answer moved from "no" to "sometimes, with a human in the loop and a verifier in the middle."
The verifier is the part builders should copy today. Aletheia's edge wasn't a bigger model; it was the natural-language checking layer that let it generate, critique and revise in a loop rather than answer in one shot. That architecture is portable.
Expect the generate-verify-revise pattern in every serious reasoning product by 2027, because raw generation without a critic is still where these systems produce confident nonsense.
My falsifiable call: by the end of 2027, at least one commercially shipped product in a regulated technical domain, most likely formal verification or computational chemistry, will cite AI-derived results as a core feature rather than a demo. The moat won't be the base model, which everyone rents. It will be the domain-specific verifier and the problem-framing scaffold wrapped around it.
Which lands founders on the harder question to sit with this quarter. If Google is openly publishing the architecture that turns a frontier model into a research agent, what are you actually selling that they can't ship as a feature next year?
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
It's a fair worry, and the raw numbers invite it: four resolved out of 700 attempted is roughly a 0.6% hit rate, per Google DeepMind's own reporting. But the significance isn't the count, it's the category. Before this, no system had autonomously closed a genuinely open research problem, and matching a known answer key is a different achievement from producing new mathematics that domain experts sign off on. The honest read is a narrow but real capability, not a solved field.
Less than you'd think on the model side, more than you'd think on the verification side. The base reasoning model is rentable from any frontier provider; the hard, valuable part is the domain-specific verifier that catches flawed candidate outputs, which is exactly what powered Aletheia rather than a larger model. Budget for the checking layer and the problem-framing scaffold, not just API calls. Expect the verifier, not the generator, to be where your engineering time and your defensibility concentrate.
Real, and it scales with how much you trust unverified output. These systems still produce confident, wrong results, which is precisely why DeepMind wraps Deep Think in a natural-language verifier and keeps a human in the loop at Level 2. In regulated or safety-critical domains, an AI-derived result is a hypothesis to be checked, not a conclusion to ship. The failure mode to fear is a plausible-looking proof or analysis that no one independently validated before it reached a customer.
