On September 21, 2026, OpenAI published a 500-word blog post that opened with one of the most consequential scientific claims of the decade and then, in the same breath, announced that a nine-person advisory panel would decide what to do about it. The claim: an internal model, trained since August 28, had resolved more than 100 long-standing open problems in mathematics in 24 days. The panel: AGMAI, the Advisory Group on Mathematics and AI, hosted at the Institute for Advanced Study in Princeton, chaired by Edward Witten, Timothy Gowers, and Martin Hairer. The mechanism by which the 100+ results will be released: deliberately undefined.
The DeepThink philosophy — transparent reasoning, inspectable chains of thought, auditable inference — was built for exactly this moment. The next twelve weeks will determine whether the field converges on closed benchmarks behind advisory gates or on open reasoning traces that anyone can replay. The stakes are not just mathematical. They are about who gets to decide what counts as a discovery.
The 100-Problem Timeline: What Actually Happened
The September 21 post is the final act of a seven-week sequence. To understand why AGMAI exists, the timeline matters:
- August 1, 2026. OpenAI announces its internal Astra model has solved ten open problems across non-sofic groups, Connes’ rigidity conjecture, sphere-packing bounds, and Erdős combinatorics. All proofs are formalized in Lean 4. Total reported compute: $2,000.
- August 28. OpenAI begins training a new internal model, beyond GPT-6 Astra. No public name; referred to internally as the “math model.”
- September 8. OpenAI drops the Navier–Stokes Millennium Prize solution. A swarm of 10,000 agents worked 88 hours, exchanged 2.7 million messages, generated ~130 billion tokens. Lean 4 formalization took 17 additional hours. The claim: finite-time blowup of 3D incompressible Navier–Stokes.
- September 9–10. Tristan Buckmaster (University of Maryland) publicly alleges his private research on singularity formation in related equations may have leaked to OpenAI. OpenAI denies. Buckmaster calls it a “Deep Blue vs. Kasparov moment.”
- September 11. 25 Fields Medalists — including Terence Tao, Peter Scholze, Maryna Viazovska, Cédric Villani, Manjul Bhargava, and 2026 medalist Yu Deng — publish an open letter titled “A Severe Misalignment of AI in Mathematics.”
- September 17. Noam Brown (OpenAI researcher) on the Dwarkesh Podcast: “I wouldn’t even attribute 10% of the credit to the multi-agent coordination. The success was the underlying model.”
- September 21. OpenAI announces AGMAI. Reveals the 100+ figure. Guest post appears simultaneously on Terence Tao’s blog.
The pattern is clear: a sequence of escalating capability claims, each one met by escalating skepticism from the mathematical community, culminating in a governance structure that places the release timeline in the hands of a panel that includes three of the same Fields Medalists who signed the rebuttal letter. AGMAI is, by design, a conflict-of-interest-bounded institution: it is the same people who complained about the pace of release being asked to bless the release itself.
The 99.9% vs. 62.7% Gap: Why Benchmark Scores Are Now Officially Not Enough
The ARC-AGI-3 numbers from September 3 deserve a second look, because they are the template AGMAI is operating under. OpenAI reported a 99.9% score on ARC-AGI-3 using its own “Provider Adapter” harness — a memory-preserving scaffolding that allows the model to maintain state across many turns. Under the ARC Prize Foundation’s standardized harness, the same model scored 62.7%. The gap of 37.2 percentage points is not a measurement error; it is the difference between “the model can do the task with our scaffolding” and “the model can do the task.”
This is the moment where the DeepThink reasoning trace becomes more than a debugging convenience. A trace is a record of every reasoning step the model took, in order, with the intermediate computations visible. If a model scores 99.9% under Provider Adapter and 62.7% under the standard harness, the reasoning trace tells you exactly where the scaffolding contributed. If a model claims to prove the Riemann Hypothesis, the reasoning trace is what the human referee reads. AGMAI’s deliberations will turn, fundamentally, on the inspection of reasoning traces — and the question of who owns those traces, who can publish them, and who can be excluded from reading them.
What AGMAI Will Probably Do, and What It Won’t
AGMAI’s structure suggests three likely decisions and one unlikely one:
Likely: AGMAI will establish a staged-release protocol for AI-generated proofs, similar to how human proofs enter the formalization queue. The first batch will be 5–10 problems; subsequent batches will follow as Lean 4 formalization clears. Expect a 60-day initial review window.
Likely: AGMAI will require that any AI-generated proof include a reasoning trace export that accompanies the publication. Whether this trace is open or gated is the central battleground. The Fields Medalists’ open letter explicitly demands transparency; OpenAI’s commercial incentives favor controlled release.
Likely: AGMAI will treat the Navier–Stokes claim with explicit skepticism, pending independent replication. The 88-hour, 130-billion-token swarm result is structurally different from anything a human mathematician has produced. Replicating it with a different model and a different formalization path is the test AGMAI will likely impose.
Unlikely: AGMAI will not block publication of results the panel finds correct. The Fields Medalists who joined AGMAI signed up for advisory work, not veto power. If a result checks out, it ships.
Why DeepThink’s Transparent Reasoning Is the Structural Antidote
The DeepSeek approach — and the DeepThink engine that powers it — has always treated the reasoning trace as a first-class artifact. Every DeepSeek API call exposes the chain of thought by default. The Engram conditional memory architecture (January 2026), the DSsec agent training sandbox (September 2026), and the V4.1-Flash CED decoder design (September 2026) all share a common design constraint: the reasoning state is separable from the output state. That separability is what makes reasoning traces exportable, replayable, and auditable.
Three concrete implications for the AGMAI debate:
- Open reasoning traces close the “closed benchmark” loophole. If a model scores 99.9% under a closed harness, the only way to verify the score is to inspect the trace. Closed benchmarks with hidden traces can be optimized; open traces with reproducible inference cannot.
- DeepThink reasoning can be replayed on different models. One of the open questions AGMAI will face is whether a result depends on the specific model weights or on the underlying reasoning. With open traces, an independent group can take the trace, run it through a different model, and check whether the conclusion follows from the reasoning or from the weights.
- Auditability creates a market for reasoning-quality scoring. Once reasoning traces are standardized artifacts, a new product category opens up: third-party reasoning-trace auditors. DeepSeek’s open-trace-by-default design makes its outputs naturally audit-ready. Closed models have to retrofit auditability, which they tend to do poorly.
The Risk: AGMAI Becomes the New Benchmark Gatekeeper
The danger in AGMAI is not its existence; it is that it becomes the only channel through which AI mathematical discoveries can be published with credibility. If the panel becomes a gate rather than a forum, two bad outcomes follow:
- Mathematics fragments. A “Mathematische Annalen” tier of peer-reviewed results emerges for human work, alongside an “AGMAI-approved” tier for AI work. Cross-pollination suffers; replication incentives weaken.
- Capture becomes structural. A nine-person panel, no matter how distinguished, is a small institution. Over time, its judgments become precedent; its membership becomes a credentialing gate. The fields that have already been through this cycle — algorithmic complexity in the 1980s, statistical learning in the 2000s — know how corrosive capture can be.
The DeepThink response to capture risk is structural, not rhetorical: publish the trace, not the verdict. If the reasoning trace is open, anyone can rerun the proof, anyone can audit the logic, and the question of whether AGMAI blessed the result becomes secondary to whether the reasoning stands on its own. This is the same logic that drove the open-source software movement in the 1990s, and it has the same structural advantages.
The Takeaway
OpenAI’s 100-math-problem claim is real, but its scope and meaning are still being negotiated. AGMAI is a reasonable institution for that negotiation, but it is not the final word. The final word will come from the reasoning traces themselves, and from the communities that can read them.
For DeepThink-powered builders and researchers, the lesson is straightforward: the next phase of AI reasoning is not about bigger models, longer contexts, or more impressive benchmark scores. It is about reasoning as an artifact. A reasoning trace that can be exported, replayed, audited, and contested is the only durable foundation for trust in AI-generated knowledge. DeepSeek has been building that foundation for nine months. OpenAI is now retrofitting it under advisory pressure. The labs that understand this asymmetry earliest will be the ones whose discoveries hold up when AGMAI-style review becomes standard practice.
The math problems will get solved. The question is who gets to verify the solutions — and whether the verification itself is open.