When an AI system questions a mathematical proof
A mathematical paper can survive expert review and still contain a mistake. Identifying it may require following a long argument and recognizing the precise assumption that fails. Google reports that Lisa Carbone, a mathematician at Rutgers University, used Gemini Deep Think to review a specialized paper in high-energy physics and infinite-dimensional algebra, and that the system identified a subtle logical flaw that had passed through human peer review unnoticed. Google’s February 2026 release post tells the story, and the Deep Think product page repeats it. It is one reported case, not proof that the system can reliably review mathematical papers in general.
This is a more consequential demonstration than a chatbot answering familiar questions. It raises a practical question for researchers and technology professionals: can specialized AI reasoning help investigate difficult problems while leaving enough evidence for experts to establish whether the answer is right?
Google’s Deep Think work offers unusually useful evidence, including abstract-reasoning results, competition-level mathematics and physics evaluations, and documented experiments in scientific research. The results also have important limits that deserve equal attention.
Deep Think is a specialized reasoning mode, not a coding-agent platform
Most AI conversations prioritize a reasonably quick response. Hard mathematics and scientific questions may require exploring several candidate explanations, testing a line of reasoning, and abandoning an answer that fails verification.
Deep Think is Google’s specialized approach to those more demanding tasks. Google’s Deep Think product page names the current offering Gemini 3.1 Deep Think and says it is built on Gemini 3.1 Pro. The same page prints the February 2026 benchmark figures under that name, while the February release post and the evaluation methodology call the same evaluation Gemini 3 Deep Think. This article reports the figures as Google reports them, with both labels, and treats them as Google’s February evaluation. They are not a measurement of every later configuration.
Google publicly describes additional inference-time reasoning work and exploration of possible solutions. It does not disclose enough implementation detail to infer the complete internal reasoning process. More computation can help explore alternatives, but it cannot guarantee a correct answer.
Antigravity is a different Google product: an agentic development environment that coordinates engineering work and tools. It is not the subject of these Deep Think reasoning evaluations.
The numbers behind Google’s reasoning claims
Google’s February 12, 2026 update reported an 84.6% ARC-AGI-2 score for Gemini 3 Deep Think. ARC-AGI-2 evaluates abstract pattern reasoning with tasks designed to challenge reliance on memorized examples. Google identifies the result as ARC Prize verified, on the semi-private evaluation set. The score is a February measurement, not the current frontier leaderboard position. Evaluation methodology
The same methodology reports 81.5% on the six official problems of the 2025 International Mathematical Olympiad, averaged across four runs and evaluated against canonical solutions using independent experts. It also reports a Codeforces Elo-equivalent rating of 3455, self-computed with a single attempt on the 222 problems of medium difficulty and above from Codeforces contests held between April and October 2025.
Scientific evaluations broaden the picture. Google reported 87.7% on the theoretical portion of the 2025 International Physics Olympiad, 82.8% on the theoretical portion of the 2025 International Chemistry Olympiad, and 50.5% on CMT-Benchmark for condensed-matter theory.
These are not five interchangeable definitions of intelligence. The Olympiad scores came from multiple evaluated runs. Physics and chemistry answers were graded against references using a Gemini judge, with Google reporting independent expert validation of gold-medal-level performance. CMT-Benchmark used pass@8 and model-based grading. The original methodology PDF makes these conditions explicit.
Accessible text alternative for this figure
ARC-AGI-2 (ARC Prize verified): 84.6%.
IMO 2025: 81.5%.
IPhO 2025 theory: 87.7%.
IChO 2025 theory: 82.8%.
CMT-Benchmark (pass@8): 50.5%.
Google reports these in its February 12, 2026 release and methodology PDF. They are Google-computed except ARC-AGI-2, which is ARC Prize verified. The Codeforces result of 3455 is an Elo-equivalent rating and is not charted.
| Evaluation | Reported result | How it was measured |
|---|---|---|
| ARC-AGI-2 | 84.6% | ARC Prize verified, semi-private set |
| IMO 2025 | 81.5% | All 6 problems, 4 runs, graded by independent experts |
| IPhO 2025 theory | 87.7% | All 3 theory problems, 8 runs, Gemini as judge, gold-medal level validated by independent experts |
| IChO 2025 theory | 82.8% | All 9 theory problems, 8 runs, Gemini as judge, gold-medal level validated by independent experts |
| CMT-Benchmark | 50.5% | Pass@8, Gemini as judge |
| Codeforces | 3455 (Elo-equivalent, not a percentage) | 222 problems, single attempt, self-computed |
A score can be impressive and still be narrower than its headline. An 87.7% physics evaluation score does not mean the system is right about 87.7% of arbitrary real-world physics research questions.
Research is harder than solving a known examination problem
An Olympiad problem has an established answer. Research may have no known solution or may depend on incomplete literature, surprising counterexamples, or experimental evidence not yet collected.
Google has investigated that harder setting using Aletheia, a research workflow powered by Deep Think. It generates candidate solutions, checks arguments, revises recoverable mistakes, and can return to the original question when a proposed solution is fundamentally unsound. The workflow can also search research literature and acknowledge when it cannot solve a problem. Google DeepMind research report
That last behavior matters. A plausible-looking incorrect proof is a poor research outcome even when the answer is fluent. A research assistant becomes more useful when its output includes a claim that specialists can independently test, not merely a persuasive explanation.
Google’s account describes work on mathematical questions, theoretical computer science, and physics, including a study of 700 open problems from a database of Erdős conjectures. In that study Aletheia autonomously solved four. Google places those results at the lowest levels of its own taxonomy of AI-assisted results and says it claims no results at its Level 3 or Level 4. On FutureMath Basic, a set of PhD-level exercises, Google reports about 38% for Deep Think and about 46% for Aletheia.
These examples are evidence of research assistance in documented cases. They are not a measured success rate for unsolved mathematics generally. Nor does a workflow that uses search and checking isolate the contribution of the underlying model from its supporting tools.
Where better reasoning helps practical science
Google also describes experiments outside competition benchmarks. Researchers have used Deep Think to investigate mathematical arguments and to develop candidate approaches in materials and physical-device design. The February release describes work with Duke University’s Wang Lab on crystal-growth conditions, which produced a recipe for thin films larger than 100 micrometers, and an engineering test by a Google Platforms and Devices lead on designing complex physical components.
These examples are promising because real scientific problems involve physical constraints, costs, measurement error, and unexpected outcomes. But a useful candidate recipe is not a universal discovery engine. Experts must verify the result in the relevant experimental environment, and the published accounts do not establish repeatable success across unrelated laboratories.
The strongest near-term use may be to narrow a search space, test reasoning, find potential mistakes, and identify promising next experiments. That is meaningful even if a human scientist remains responsible for the decision.
Why extra reasoning does not eliminate uncertainty
Inference-time computation is the processing performed while a trained model works on a new request. Specialized reasoning approaches may allocate more of it to difficult problems, permitting additional exploration and checking.
The trade-off is that more processing can increase latency and computational cost. It may also produce additional incorrect candidate arguments. Google’s public materials do not establish a fixed relationship between more thinking time and the probability of a correct answer on every task.
Deep Think should therefore be judged not only by the quality of a final answer, but also by how that answer can be checked. Mathematics may offer a formal proof or a reference solution. Materials research may require an experiment. A decision based on incomplete business data may need a different kind of validation.
The appropriate verification method depends on the problem. There is no single benchmark percentage that substitutes for it. TechiesJournal’s guide to how to read AI model claims covers the general habit of asking what a benchmark number actually measured.
Access to Deep Think
At the February 2026 upgrade, Google stated that the updated Deep Think mode was available in the Gemini app for Google AI Ultra subscribers, with an interest-based early-access route for selected researchers, engineers, and enterprises through the Gemini API. These were the stated access conditions at release, not a guarantee about every plan, region, or API option today. Google’s release announcement Google’s current product page offers a “Try in Gemini” button and states no plan, price, or API conditions.
A team evaluating the system should check its own current availability and terms. It should also ask whether the problem really benefits from the specialized mode. Routine summarization and straightforward coding are not necessarily good reasons to spend more time and computation on extended reasoning.
Deep Think within Google’s evolving model family
The dedicated Deep Think materials focus on a specialized reasoning offering, but Google’s wider Gemini family has continued to advance. On November 10, 2026, Google announced Gemini 4 Carbon as its latest frontier model. That does not retroactively turn the February Deep Think measurements into Carbon measurements or establish that the named Deep Think mode and Carbon are the same product. Google’s Carbon announcement
The practical distinction is between a specialized reasoning mode and the newer general frontier-model lineup. Both matter when assessing Google’s progress, but they answer different product questions.
How Google compares with Anthropic and OpenAI
Google’s February benchmark material compared Deep Think with Claude Opus 4.6 at maximum thinking and GPT-5.2 at its highest reasoning setting, as shown in the product page table and the methodology PDF. Those historical comparisons are useful context, not the current leaderboard.
For a more recent single-benchmark comparison, the ARC Prize results table lists the dated ARC-AGI-2 outcomes below. The first row is Google’s own February figure, which Google reports as ARC Prize verified, and it does not appear in the table today. The other rows are ARC Prize listings.
Accessible text alternative for this figure
Gemini 3 Deep Think (Feb 12): 84.6%.
GPT-6 family (Sep 2): 95.0%.
GPT-6.1 Sol (Sep 29): 94.2%.
Gemini 4 Argon (Oct 13): 91.3%.
Claude Fable 5.2 (Nov 9): 93.8%.
Gemini 4 Carbon (Nov 10): 94.2%.
The Gemini 3 Deep Think figure is Google’s own February announcement, ARC Prize verified per Google. The other rows are from the ARC Prize results page reviewed on November 11, 2026. The GPT-6 row is a family listing for Astra, Sol and Luna. A single benchmark does not rank overall ability.
| Provider and evaluated model | ARC-AGI-2 score and date |
|---|---|
| Google, Gemini 3 Deep Think | 84.6% on February 12, 2026 (announced) |
| OpenAI, GPT-6 family listing | 95.0% on September 2, 2026 |
| OpenAI, GPT-6.1 Sol | 94.2% on September 29, 2026 |
| Google, Gemini 4 Argon | 91.3% on October 13, 2026 |
| Anthropic, Claude Fable 5.2 | 93.8% on November 9, 2026 |
| Google, Gemini 4 Carbon | 94.2% on November 10, 2026 |
The table is a selection of recent rows, and the ARC Prize page lists other models. These are dated results from particular evaluations. ARC Prize lists GPT-6 as one family row covering Astra, Sol, and Luna, so that score should not be attributed to one individual variant. The four highest scores in the table, 93.8% to 95.0%, fall within 1.2 percentage points of each other, and in TechiesJournal’s analysis a difference of less than one percentage point should not be treated as proof that one model is more intelligent across all subjects.
Google: The February Deep Think result was a notable reasoning milestone. ARC Prize lists Gemini 4 Argon at 91.3% on October 13 and Gemini 4 Carbon at 94.2% on November 10, so the broader Gemini family has moved past the February figure on this measure.
Anthropic: Claude Fable 5.2 is positioned for demanding technical work, including science and engineering. Anthropic’s November announcement reports strong performance on coding and scientific tasks and does not mention ARC-AGI-2, so the 93.8% comes from the ARC Prize listing. Those announcement results are not direct substitutes for Deep Think’s Olympiad and condensed-matter evaluations. TechiesJournal’s earlier article on Fable 5.2 and Mythos 5.2 covers them.
OpenAI: ARC Prize lists the GPT-6 family and GPT-6.1 Sol with high ARC-AGI-2 scores. OpenAI’s model documentation names the variants (gpt-6-astra, gpt-6.1-sol, and gpt-6-luna) and does not report ARC-AGI-2 figures, so variants and effort settings must be specified when comparing costs, latency, or performance. TechiesJournal’s earlier comparison of GPT-6 Astra and Claude Fable 5.1 shows how the two vendors’ earlier models stacked up.
The strongest practical comparison is not a universal winner. A laboratory, developer, or researcher should evaluate the models on representative problems, require independently checkable outputs, document cost and latency, and compare failure patterns under equivalent conditions.
The meaningful advance
Gemini Deep Think shows that AI reasoning can contribute to work beyond ordinary question answering. The reported mathematics and scientific evaluations demonstrate strong performance on selected demanding problems, while the research collaborations show how experts are beginning to use these systems to investigate real questions.
The remaining challenge is verification. A model that finds an overlooked flaw or suggests a successful experiment can be valuable, but its answer must survive the evidence standard of the field in which it is used.
For researchers and engineering teams, the useful next step is to select one difficult, verifiable problem and assess whether specialized reasoning improves the quality of the result enough to justify its cost and review effort.
AI reasoning becomes genuinely useful when it helps people reach conclusions they can independently establish as sound.
Go deeper
Sources were checked on November 11, 2026.
- Google DeepMind: Gemini Deep Think: Current product overview, benchmark table, and research examples.
- Google: Gemini 3 Deep Think upgrade: February 2026 release, access description, and the Carbone and Wang Lab examples.
- Google Deep Think evaluation methodology (PDF): Benchmark configurations and grading qualifications.
- Google DeepMind research report: Aletheia and the research evaluations.
- ARC Prize results: Dated independent benchmark records.
- Google Gemini 4 Carbon release: November 2026 model-family context.
- Anthropic Claude Fable 5.2 and Mythos 5.2: Current vendor claims and availability.
- OpenAI API model catalog: Current model identifiers and access information.