Deep Think Mathematica: Inside Google's Leaked Math Model, and the Scream in Its Reasoning Trace

The most revealing thing about Deep Think Mathematica — the Goo​gle math model that surfaced in internal testing this week — is that it appears to shout. An X account posting under the name "Lyra" put screenshots of internal reasoning logs online on September 16, 2026, and the traces read less like a proof assistant than a person having a breakthrough at a whiteboard: "Wait." "Oh my god." And, after simplifying an equation successfully, the string HOLY MOTHER OF MATHEMATICS!!!!!!!!!!!!!!!!!!!!!!!! The reaction was immediate and affectionate — Goo​gle had apparently built an AI that genuinely loves math. That reaction is also the least useful thing to take from the leak. Nothing here has been confirmed by Goo​gle. There is no model card, no model ID in any published list, no price, and no release date. What exists is a build string, two internal labels, a token budget, a set of screenshots, and the best shipping comparison point in the same family, Gem​ini 3.1 Deep Think — a model you can actually use today, and the yardstick this leak will eventually be judged against.

So treat everything below as what it is: a claim list with a source attached to each item, sorted by how much weight it will bear.

What the signal actually says — and the tell sitting above it

The signal itself came from @testingcatalog, an account with a decent record on Goo​gle specifically: it surfaced the Gem​ini desktop computer-use plans, and it caught Nano Banana 2's early preview card under the internal name "GEMPIX2" before Goo​gle said anything. Its September 16 post carries two claims of very different quality.

The second claim is the model. A new "Deep Think Mathematica" has been spotted in internal testing, described as a successor in spirit to the Deep Think IMO model Goo​gle shipped to selected institutions last year.

The first claim is the tell, and it is the one worth understanding: "When Gem​ini is about to change its background, something big is brewing." That is not a fact about a model. It is a community heuristic about Goo​gle's release choreography — the observation that a visible refresh of the Gem​ini app's surface has preceded several of Goo​gle's larger announcements. Heuristics like this have a real track record and zero predictive force. A UI refresh is not a model.

Both claims are unverified. Neither is a release, and this piece does not treat it as one.

Decoding the build string, because it is the most concrete artefact here

The single hardest piece of evidence in circulation is a build identifier attributed to the leaked logs:

• Build path — models/deepthink-mathematica-tf-raw-thoughts, reported by AI Sight on September 16, 2026, alongside the screenshots.

That string rewards reading closely, and it is where most coverage stops. Three parts of it carry information.

• "deepthink" — places the model in the Deep Think line rather than the mainline Gem​ini line. That matters because Deep Think is a mode in Goo​gle's shipped products, not a separate family: it is the parallel-reasoning path that runs multiple candidate solutions at once.

• "tf" — in context, an abbreviation for "Teamfood," the second of the two internal labels reported alongside the build. In Goo​gle's internal vocabulary that is an early, dogfooding-stage designation: employees using it, not customers.

• "raw-thoughts" — the most interesting part, and the key to the shouting. This denotes the unfiltered chain-of-thought configuration — the model's reasoning emitted without politeness tuning, style constraints or output filtering. A "raw thoughts" build is not the product. It is the thing engineers look at when they want to see what the model is doing before anyone makes it presentable.

The second reported label is UNSTABLE_EXPERIMENTAL, which is close to self-explanatory and points the same direction as "Teamfood": early, internal, not for customers.

Two further figures are attributed to the same logs and are single-source, so hold them loosely:

• Context window — roughly 1,000,000 tokens, matching the 1M window Goo​gle already ships on its frontier Gem​ini models.

• Output limit — 65,536 tokens, which for a reasoning model means a very long deliberation before an answer.

That is the entire technical surface of the leak. A build path, two stage labels, a context window and an output cap. It is enough to say a thing exists in a test harness. It is not enough to say what it scores.

Why a reasoning trace that yells is weaker evidence than it looks

The exclamation marks are the reason this leak travelled, and they are the reason to be careful with it.

The reported explanation is mundane and fits the build string exactly: an unfiltered reasoning configuration, run at high generation temperature, with the politeness and formatting layers stripped out. When you remove the tuning that makes a model sound calm, you do not reveal its soul — you reveal its sampler. Exclamation-heavy output is what a high-temperature sample of an excited continuation looks like. The emotional reading is an artefact of the configuration, not a finding about the model.

The deeper point is about what a chain of thought is. A reasoning trace is generated text that happens to be useful for solving problems. Reading a transcript of it and inferring an internal state — enthusiasm, frustration, delight — is the same category error as inferring that a calculator is pleased when it produces a clean integer. It is worth stating plainly because the Chinese-language coverage of this leak leaned into the joke ("等等", "我的天") and some English coverage let the joke harden into a description of the model's character.

None of that makes the trace worthless. Exposing a raw-thoughts build is genuine evidence about a real engineering practice — you only log unfiltered reasoning when you are debugging the reasoning itself, which is what you do with a model whose entire value proposition is how well it thinks through hard problems. And it is unverified whether the traces came from deliberate debugging or unintended logging.

The honest summary: the exclamations tell you a raw reasoning configuration exists. They tell you nothing about mathematical capability.

Millennium Bench is not the Millennium Prize Problems

Several write-ups of this leak, including the aggregator summaries circulating on September 16, have the model "targeting Millennium Bench" and leave it there. The name is doing accidental work, because it sits one word away from something much more famous and much more loaded — the Clay Mathematics Institute's seven Millennium Prize Problems, each carrying a $1 million bounty, and the subject of a separate and entirely unconfirmed claim this month about Ope​nAI and a Navier–Stokes proof. One report from that thread says Goo​gle built an AI that "solved" the problem in 88 hours. Clay still lists it open.

These are different things, and conflating them inflates this leak considerably.

MILLENNIUM-BENCH, as introduced in the ICLR 2026 paper "Where Deduction Breaks Down: Diagnosing Reasoning Failures in Research-Level Mathematics," is a benchmark of expert-curated research-level problems spanning nine domains, including mathematical physics, topology and measure-theoretic probability. It was built to require sustained multi-step deduction well past competition mathematics.

Its headline result is the part that matters for reading this leak. Across five frontier models, run 100 times per problem, the strongest achieved at most 24% pass@1 and 39% pass@3. The paper's diagnosis of how models fail is more useful than the scores: framework hallucination, where a model invents an inapplicable theory and then reasons coherently inside it, and fabricated theorems, where it cites results that do not exist, complete with invented author names and theorem numbers.

Hold both of those in mind together. A benchmark where the best model tops out near a quarter of problems, whose signature failure is confident invention, is exactly the benchmark where leaked screenshots are least able to demonstrate anything. A model that shouts while it works is a model whose output you would want scored by someone other than its author.

What Goo​gle has actually shipped in this line

The leak is only legible against the shipped record, because Goo​gle's math story is a documented progression and the leaked name is borrowing its credibility.

• July 2025 — an advanced version of Gem​ini Deep Think reached gold-medal standard at the International Mathematical Olympiad, solving five of six problems for 35 of 42 points, graded by IMO officials under the same standards applied to students. Demis Hassabis noted it worked end-to-end in natural language without translation into formal syntax.

• August 1, 2025 — Goo​gle shipped Gem​ini 2.5 Deep Think publicly, but not the gold model. The released version was described by Goo​gle as a faster, more optimized variation, reaching Bronze-level performance on the 2025 IMO benchmark per Goo​gle's internal evaluation. It was gated to the top consumer tier. Simultaneously Goo​gle gave the full gold model to a select group of mathematicians and academics — the "selected institutions" the leaked signal is referring back to.

• December 2025 — Gem​ini 3 Deep Think, with a large generational jump Goo​gle reported as 45.1% on ARC-AGI-2 with code execution and 41.0% on Humanity's Last Exam without tools.

• Now — Gem​ini 3.1 Deep Think, built on Gemini 3.1 Pro, sold through the top consumer subscription tier and an invitation-only API programme. Goo​gle's own published figures, which are vendor-reported and have not been independently reproduced, put it at ARC-AGI-2 84.6%, Humanity's Last Exam 48.4% without tools and 53.4% with search and code, IMO 2025 81.5%, and a Codeforces Elo of 3455.

Two adjacent efforts matter too. Aletheia, a math research agent built on Gem​ini Deep Think, is reported by third parties at roughly 95.1% on IMO-ProofBench Advanced — and is notable less for the number than for being designed to say "no solution found" rather than invent one; on the February 2026 FirstProof challenge it solved 6 of 10 and declined the rest. And an AI Co-Mathematician setup running on Gemini 3.1 Pro reportedly lifted FrontierMath Tier 4 performance from 19% to 47.9%.

That last figure is worth pausing on, because it is the strongest argument against reading this leak the way it was written up. A jump from 19% to 47.9% on research-level mathematics, achieved by a scaffold wrapped around a general Gem​ini model rather than a new checkpoint, suggests the largest recent gains in Goo​gle's math performance have come from the harness around the model, not from a specialist model underneath it. It does not mean Mathematica is a scaffold. It does mean the burden of proof for "this is a new dedicated math model" is higher than the coverage assumed.

One more piece of context sits over all of it: DeepMind restructured in August 2026, with Demis Hassabis moving into a chairman role. Leaks from inside a reorganizing lab are not more reliable than leaks from a stable one, and the coverage has not treated the timing as relevant. It might be.

Model or scaffold? The question this leak does not settle

There is a live argument in the coverage about whether Mathematica is a new model at all. One camp points at the build string's "deepthink" prefix and reads it as a separate checkpoint. Another camp — which our own earlier write-up on the Deep Think V3 rumour landed in — holds that Goo​gle's specialist math results come from agent scaffolds wrapped around general Gem​ini models, and that no separate math model needs to exist for the numbers to be real.

The build string is a modest point in the first camp's favour: models/deepthink-mathematica-tf-raw-thoughts names a model, and "raw-thoughts" describes a model configuration rather than a harness layer. A scaffold would not need its reasoning unfiltered to be debugged; a model whose thinking is the product would. That is a real signal.

It is not a decisive one. Harnesses have configuration too, and an internal path string is written by whoever set up the logging, not by a schema. What would settle it is the thing nobody has: a model card, or an endpoint, or a benchmark run by someone without a stake in the answer.

What you can actually call today

None of this changes anything you can put in production this week, and it is worth being blunt about the gap. Deep Think mathematica is not on any API. Gem​ini 3.1 Deep Think is not sold per token either — it is gated behind the top consumer subscription tier, listed between $99.99 and $199.99 a month depending on tier, plus an invitation-only API programme. There is no published per-million-token price for it.

The base model it reportedly builds on is a different story. Gemini 3.1 Pro is priced at $2.00 per million input tokens, $0.20 cached, and $12.00 per million output tokens for contexts under 200,000 tokens, rising to $4.00 / $0.40 / $18.00 above that — Goo​gle's own published rates.

That pricing detail is where the practical decision lives, because the cost gap between reasoning modes is not small. On ARC-AGI-2, Deep Think is reported at roughly 85% for about $13.62 per task, against Gemini 3.1 Pro at 77% for about $0.96 per task. You are paying roughly fourteen times more for eight points. That is a defensible trade for a hard research problem and an indefensible one for a production path that mostly handles ordinary work — and the correct answer is usually that you want both reachable from the same place rather than a subscription decision made in advance.

That is the case for routing on one key. The Gem​ini models that are callable today — Gemini 3.1 Pro Preview, Gemini 3.8 Flash at $0.750 per million input tokens with a $0.075 cache read, and the rest of the family — sit on OrcaRouter's catalogue of 198 models across 15 providers, behind one OpenAI-compatible endpoint. There is no markup on provider list pricing, so a Goo​gle price change is live on our side the same day rather than waiting on a re-contract. Automatic failover means an unproven or preview model can sit in a route without carrying a production path on its own, which is exactly the posture a leaked model deserves: try it, keep a fallback, change nothing else. The routing DSL composes several models into one call, and model fusion runs a panel and combines the answers — both useful when the question is a hard one and the honest position is that you want more than one model's opinion.

To be explicit about the boundary: OrcaRouter does not host Deep Think mathematica, and it does not host Gem​ini 3.1 Deep Think. Those are not on our catalogue and nothing here should suggest otherwise. What is on the catalogue is the Gem​ini layer underneath them.

What would turn this from a leak into news

Four things, in rough order of how much they would move the story.

• A model card. Goo​gle's Deep Think releases have come with published evaluations. A card with numbers on MILLENNIUM-BENCH or IMO-ProofBench Advanced, run under stated conditions, would convert this from a build string into a model.

• An endpoint. The clearest signal would be Mathematica appearing in the Gem​ini API or in Goo​gle's model list under any name. Until that happens, no one can call it and no price exists to compare.

• The institution route. Goo​gle's pattern in 2025 was to give the strongest math model to selected mathematicians first and the public a faster variant later. A quiet rollout to researchers would fit that precedent and would likely surface before any product announcement.

• A third-party run. Given that the benchmark in question tops out near 24% pass@1 for the best available models, and its signature failure mode is confident fabrication, an independent evaluation is the only evidence that will hold. Every figure in this piece that carries a number is either Goo​gle's own, third-party reported, or explicitly flagged as unverified — and none of them came from a model anyone outside DeepMind has run.

The one thing worth doing now is not waiting. The gap between Goo​gle's math mode and its base model is fourteen-fold in cost and eight points in accuracy, and that trade is already live; you can make it today with Gemini 3.1 Pro Preview on one key and a fallback behind it. When Mathematica gets a model card, you will want to have already decided how you route hard problems — and the answer to that will not depend on what the leaked model turns out to be.