WorldbyFlow ForecastsOpen · the question and its record →
Make this research yours. Add it to a free WorldbyFlow workbench to run follow-ups, ask questions, and re-check it as events move.
Add to your workbench — free
WorldbyFlow•Structured Research
Generated September 28, 2026· 40 sources

AI lab solving another Millennium Prize Problem before 2027

Forecast
Share
Our forecast
8%
Kalshi market
43%
35 points below the market
The probability this question resolves YES under its own rules, computed from the chain below: an outside-view starting point of 3%, then each graded adjustment added in log-odds. The market price is shown for reference; it is not an input to the chain.
Why it leans this way
A second AI-produced Millennium proof announcement inside the window is genuinely plausible, but this market pays on peer-reviewed publication, official recognition, or two Source Agencies reporting the problem resolved by reference to a written work — and the Navier-Stokes precedent showed that even the loudest possible announcement produced coverage framing it as an unadjudicated company claim.

How the Number Is Built

Starting point3%
Millennium Prize Problems achieving qualifying recognition (peer-reviewed publication of a complete solution, or official CMI/governing-body statement of resolution) per unit time, across the 26 years since the problems were posed in 2000: 1 of 7 problems has ever been officially declared solved (the Poincaré conjecture, prize awarded 2010), across 26 years. The chosen base rate, converted to the time this question has left.
▲Demonstrated capability to produce a Millennium-scale proof artifact in days3% → 19% +16
decisive toward yes · documented
▲A named second-problem effort reported as near completion19% → 38% +19
strong toward yes · reported
▼The settlement standard excludes exactly the artifacts a lab produces first38% → 19% −19
strong toward no · documented
▼The 20-day Navier-Stokes record shows announcement does not convert to recognition on this timescale19% → 12% −7
moderate toward no · documented
▼No written work exists yet for any second problem12% → 8% −4
moderate toward no · documented
▼Reported intent to delay the announcement for messaging reasons8% → 6% −2
slight toward no · reported
▲A second lab independently active on the remaining problems6% → 8% +2
slight toward yes · reported
Our forecastKalshi market 43%, shown for reference, never a step8%
Each size is a fixed step in the odds (slight 0.2, moderate 0.5, strong 1.0, decisive 2.0), so the same step moves the number more near 50% than near 0% or 100%. Steps add up, so the order they are listed in does not change where the forecast lands.

The Question

An AI lab must solve a Millennium Prize Problem other than the already-claimed Navier-Stokes, after Issuance and before January 1, 2027, where "solved" means the first unretracted date on which either a complete proof, disproof, or counterexample addressing the problem as originally formulated by CMI is published in a reputable peer-reviewed journal; or CMI or another recognized academic governing or recognition body officially states the problem has been resolved by reference to a specific written work; or at least two Source Agencies themselves report the problem proven, disproven, or otherwise resolved by reference to a specific publicly available written work. The lab, its personnel, or an AI system it developed and operated must be credited as author, contributor, or producer.
Deadline: 2027-01-01T04:59:00Z, per the market's stated close
Settled by: The Guardian, the National Academy of Sciences of the United States, the Financial Times, the Associated Press, and the Clay Mathematics Institute
Kind of question: occurrence
Ambiguity in the rules: material — Two genuine ambiguities. First, whether Navier-Stokes itself is excluded: the question says "another" problem and "after Issuance," and the rules say problems solved before the time period do not count, but the Issuance date is not stated in this record and CMI's September 11 "apparently been settled" statement post-dates the September 8 announcement — if Issuance predates September 11 and that statement were read as an official recognition by reference to a specific written work, a settlement dispute is conceivable. The sibling structure and the adjacent Kalshi contract excluding Navier-Stokes by name both point to exclusion, which is the reading used here. Second, the threshold for a Source Agency "reporting the problem resolved" versus reporting a lab's claim is a judgment call that the rules do not further define.
Where the rules decide it
  • A preprint, announcement, blog post, conference presentation, technical report, or machine-checked Lean formalization does NOT by itself constitute a solution — so a repeat of the September 8 Navier-Stokes playbook, however loudly covered, does not resolve YES on its own. This cuts strongly toward NO and is the single most important rule in the market.
  • The two-Source-Agency pathway requires the agencies to report that the problem has been RESOLVED by reference to a specific written work. Coverage that frames a result as a company claim, as Nature did on September 8, does not obviously satisfy this, and the settlement source list is narrow — the Guardian, AP, FT, NAS, and CMI. Quanta, Nature, the NYT, Scientific American, and WIRED coverage does not count toward the two-agency threshold.
  • Partial results, special cases, unrelated reformulations, and results conditional on unproven assumptions do not count; the result must establish truth or falsity of the problem as originally formulated. The Navier-Stokes claim addressed the forced version (Clay alternatives C and D) while the unforced global-regularity question remains open, illustrating exactly how a headline-grade result can fall short of the original formulation.
  • A purported solution retracted, withdrawn, repudiated, or reported to contain an unresolved identified error before Expiration does not count — so a late-December announcement carries retraction risk inside the window, and the live Navier-Stokes priority dispute shows how fast contestation arrives.

The Forecast

This market pays only if a Millennium Prize Problem beyond Navier-Stokes is solved by an AI lab after Issuance and before January 1, 2027, and — critically — only on a peer-reviewed publication, an official recognition body statement, or two Source Agencies reporting resolution by reference to a specific written work. A lab announcement with a preprint and a Lean file, which is exactly what the Navier-Stokes precedent produced, does not by itself satisfy the rules. The evidence points to a live but unconfirmed second effort on the Hodge Conjecture with no paper, no code, and no institutional statement as of the latest reporting, against a 94-day window and a settlement standard that normally takes far longer than that to clear.
The question looks like a capability forecast and is actually a recognition-process forecast. Under the verbatim rules, a Millennium Prize Problem counts as solved only on the first unretracted date that one of three things happens: publication of a complete proof, disproof, or counterexample in a reputable peer-reviewed journal addressing the problem as originally formulated by the Clay Mathematics Institute; an official statement from CMI or another recognized academic governing body that the problem has been resolved by reference to a specific written work; or at least two Source Agencies themselves reporting the problem proven or resolved by reference to a specific publicly available written work. The rules then explicitly strip out the artifacts an AI lab actually produces first: a preprint, an announcement, a blog post, a conference presentation, a technical report, and a machine-checked formalization do NOT by themselves constitute a solution.
Where things stand is unusually well documented for a 94-day window. On September 8, 2026, OpenAI published a claimed resolution of Navier-Stokes produced by an internal model, with a 166-page manuscript and a Lean formalization, and the surrounding record is consistent across Tier 1 outlets. That problem, however, is almost certainly out of scope here: the market asks about "another" problem solved "after Issuance," the sibling structure and the adjacent Kalshi contract on the same theme exclude Navier-Stokes by name, and the rules say problems solved before the time period do not count. The forecast therefore rests on a second problem clearing the recognition bar inside the window.
The second track is real but thin. The Information reported, citing a single OpenAI source, that employees expect to crack the Hodge Conjecture, and OpenAI separately told the New York Times it had made "substantial progress" on an unnamed second Millennium problem. Against that, coverage through late September records no paper, no code, and no statement from any authoritative institution on Hodge, and the same reporting notes OpenAI may deliberately delay any announcement while it works out how to present it without repeating the Navier-Stokes backlash. The Birch and Swinnerton-Dyer chatter traced to a single X user's explicitly labeled prediction rather than to any lab.
The outside view is harsh on this window. The recognition machinery for Millennium problems runs on a multi-year clock: CMI's own rules require publication in a qualifying outlet, at least two years elapsed, and general acceptance in the mathematical community, and its September 11 statement on Navier-Stokes said only that the problem had "apparently been settled" while describing its evaluation as deliberately unhurried. Peer review in mathematics for a result of this magnitude does not complete in three months. That leaves the two-Source-Agency pathway as the realistic route to YES, and it is a genuinely plausible one, because a sufficiently clean announcement covered by the Guardian, AP, and the FT could produce two agencies reporting a problem resolved by reference to a specific written work. That pathway is also where the rules bite hardest, since the 2026 record shows major outlets consistently framing the Navier-Stokes result as a company claim rather than an adjudicated solution.
The chain builds from a base rate for a qualifying recognition event in a 94-day window, lifts substantially for the documented capability step-change and the reported second-problem effort, and then takes back weight for the settlement standard and the reported intent to delay. The result sits somewhat below the market's roughly 43%, and the market's own path — 68¢ on September 13 down to 40¢ today — suggests traders have been migrating in the same direction as the rules-based read.
Latest evidence: 2026-09-28

If Nothing Changes

Resolves NO if nothing changes
If nothing changes, the market resolves NO. As of the latest coverage there is no written work in existence for any second problem — no named paper, no Lean artifact, no lab confirmation on Hodge or BSD — and CMI lists the remaining problems as open. No peer-reviewed publication is pending, and no recognition body has stated any second problem resolved.
Time left: 94.4 days. For YES, a second problem would need a written work produced, published or announced, and then reported as resolved by two of five narrowly specified Source Agencies, all inside roughly three months. The September 8 precedent shows the produce-and-announce phase can run in days once a model is turned loose, which is why the window is not prohibitive — but the recognition phase has no precedent of completing this fast.

The Outside View (3)

How often comparable situations resolved this way, before the specifics. “Counted” means the cases are named and counted; “estimated” means no count was available and the rate is an estimate, said so.
Millennium Prize Problems achieving qualifying recognition (peer-reviewed publication of a complete solution, or official CMI/governing-body statement of resolution) per unit time, across the 26 years since the problems were posed in 2000The base rate usedcounted
Base rate: 1 of 7 problems has ever been officially declared solved (the Poincaré conjecture, prize awarded 2010), across 26 years · 2%
Cases: Poincaré conjecture — Perelman preprints 2002-2003, CMI prize awarded 2010 — is the sole officially resolved case. The other six (Riemann, P vs NP, Yang-Mills, Navier-Stokes, Hodge, Birch and Swinnerton-Dyer) had no qualifying recognition event through the start of this window.
Fit: This is the correct denominator for the recognition event the rules require, and it captures the multi-year institutional clock. It badly breaks on the numerator side, because it is drawn almost entirely from a pre-AI era in which no actor could generate a candidate proof in days — which is what the adjustments exist to correct.
Source: Counted from the cases listed; CMI and Wikipedia both record Poincaré as the only officially declared solution.
Reported-imminent AI lab Millennium-problem results converting to a qualifying recognition event within 90 days of the report, during the September 2026 episodeestimated
Base rate: 0 of 2 reported near-solutions (Hodge via The Information; BSD via social media) has produced a paper, code, or institutional confirmation as of late September 2026 · 15%
Cases: The Hodge Conjecture report (The Information, single OpenAI source, September 17) and the BSD chatter (traced to one X user's labeled prediction) are the only two instances; neither had converted to a written work as of the latest coverage.
Fit: Directly on-point for the actual causal question, but the sample is two open cases, so it cannot bear weight as a prior. Useful mainly as a sanity check that nothing has converted yet.
Source: Estimated — the episode is too new and the sample too small (n=2, both unresolved) to count a rate. Stated as estimated rather than dressed up as a base rate.
The Navier-Stokes precedent: time from AI lab announcement to a qualifying recognition event under these rulescounted
Base rate: 0 of 1 — 20 days elapsed from the September 8 announcement to September 28 with no peer-reviewed publication and no official statement of resolution · 10%
Cases: OpenAI Navier-Stokes, announced September 8, 2026: manuscript and Lean formalization public; CMI's September 11 statement said "apparently been settled" with evaluation "deliberately unhurried"; Nature framed it as a company claim; no journal publication.
Fit: The single most informative case, because it is the same labs, the same rules environment, and maximum media attention. It shows the announce-to-recognition gap is long even in the best case. n=1 limits it to a corroborating class rather than the prior.
Source: Counted from the single case; CMI statement and Nature coverage in grounding.
For this window: Taking the counted historical class: 1 qualifying recognition event in 26 years is roughly 0.038 per year. For a 94.4-day window, the chance of at least one event is 1 - (1 - 0.038)^(94.4/365) = 1 - 0.962^0.259 ≈ 0.010, about 1%. That is the honest pre-AI outside view and is plainly too low for the current period, so I set the prior slightly above it at 3% to reflect that the window also contains a second lab (Anthropic) and a broader set of active efforts than the historical average year, while leaving the capability step-change entirely to the adjustments where it can be graded.
Starting point: 3%

What Moves It (7)

Specific evidence about this question that the starting point does not already carry. A claim’s weight can never exceed its grade: an inferred adjustment counts at most moderate, a contested one at most slight.
Demonstrated capability to produce a Millennium-scale proof artifact in daysdocumented
toward yes · decisive
Evidence: On September 8, 2026, OpenAI published a 166-page manuscript and Lean formalization for a Navier-Stokes finite-time singularity, produced in roughly 88 hours by about 10,000 concurrent agents on an internal model that had been in training only since late August 2026 and scored close to 50% on the company's internal open-problem benchmark versus about 10% for GPT-6 Astra.
Not already counted because: The 26-year base rate is drawn from a period in which no actor could generate a candidate Millennium proof in days, so the reference class cannot contain this mechanism at all.
Source: OpenAI announcement and Quanta Magazine, September 8, 2026; benchmark figures per OpenAI researcher quoted in September 2026 coverage
A named second-problem effort reported as near completionreported
toward yes · strong
Evidence: The Information reported on September 17, 2026, citing a single source at OpenAI, that employees expect to soon crack the Hodge Conjecture; OpenAI separately confirmed to the New York Times that it had made "substantial progress" on another Millennium Prize problem without naming which.
Not already counted because: The historical class contains no instance of an actor publicly signalling an imminent second Millennium solution; this is a specific, dated signal about this window.
Source: The Information, September 17, 2026; OpenAI statement to the New York Times as reported in September 2026 coverage
The settlement standard excludes exactly the artifacts a lab produces firstdocumented
toward no · strong
Evidence: The market's additional rules state verbatim that a preprint, announcement, blog post, conference presentation, technical report, or machine-checked formalization does NOT by itself constitute a solution, and require either peer-reviewed publication, an official recognition-body statement by reference to a specific written work, or two Source Agencies reporting the problem resolved by reference to a specific written work.
Not already counted because: The base rate measures recognition events and so partly embeds this, but the reference class cannot capture that this specific market's settlement path is narrower than CMI's own — it excludes the Lean formalization that is the labs' main credibility instrument.
Source: MARKET RECORD, additional rules, verbatim
The 20-day Navier-Stokes record shows announcement does not convert to recognition on this timescaledocumented
toward no · moderate
Evidence: Twenty days after the September 8 announcement, there is no peer-reviewed publication; CMI's September 11 statement said only that the problem "has apparently been settled" and described its evaluation as "deliberately unhurried"; Nature treated the result as a company claim; Wikipedia records it as unverified by CMI or the independent mathematical community and subject to a priority dispute.
Not already counted because: This is a new, same-regime observation post-dating the reference class entirely, and it is specifically about conversion speed under maximum attention rather than about base incidence.
Source: CMI statement, September 11, 2026; Nature, September 8, 2026; Wikipedia Clay Mathematics Institute entry as of late September 2026
No written work exists yet for any second problemdocumented
toward no · moderate
Evidence: Coverage through late September 2026 records no named paper, no code, and no statement from any authoritative institution on the Hodge Conjecture; CMI continues to list the remaining problems as open; the Birch and Swinnerton-Dyer claim traced back to a single X user's explicitly stated prediction rather than to any announcement from Anthropic or OpenAI.
Not already counted because: The reported-effort adjustment above captures the intent signal; this captures the separate fact that the clock on any recognition pathway has not started, which is independent of whether the effort succeeds.
Source: Fact-checking coverage and Clay Mathematics Institute problem listings, September 2026
Reported intent to delay the announcement for messaging reasonsreported
toward no · slight
Evidence: The Information's source said it could take the company longer to announce a solution because it is trying to figure out how to collaborate with the math community to make the announcement without triggering another public-relations problem, following the September backlash in which several mathematicians characterized the Navier-Stokes episode as research misconduct.
Not already counted because: A deliberate announcement delay is a mechanism specific to this window and this actor, created by the September backlash; no historical case involved a solver with a reputational reason to hold back.
Source: The Information, September 17, 2026, as paraphrased in Gizmodo and Decoder coverage, September 2026
A second lab independently active on the remaining problemsreported
toward yes · slight
Evidence: Anthropic reported on September 4, 2026 that Claude had formalized an existing proof of Fermat's Last Theorem using about 29,500 intermediate theorems, and September 2026 commentary from someone describing knowledge of mathematicians at both labs reported that a substantial fraction of 2026 mathematical progress is known only to OpenAI and Anthropic insiders.
Not already counted because: The base rate implicitly assumes one-solver-at-a-time academic effort; two well-resourced labs racing raises the arrival rate independently of either one's specific reported progress.
Source: September 2026 coverage of Anthropic's Fermat formalization; LessWrong commentary, September 13, 2026

Against the Market

Kalshi market: 43% · our forecast 8% · 35 points below the market
The market at roughly 43% sits above this chain. The most likely explanation is that traders are pricing the probability of a second AI Millennium ANNOUNCEMENT rather than the probability of a qualifying recognition event under these specific rules — the two diverge sharply here, because the rules exclude preprints, announcements, and Lean formalizations on their own, and restrict the two-agency pathway to a narrow list of the Guardian, AP, FT, NAS, and CMI. An announcement is genuinely likely inside 94 days given the documented capability and the reported Hodge effort; two of those five specific agencies reporting a problem RESOLVED by reference to a written work is a distinctly higher bar, and the Navier-Stokes precedent produced coverage framing the result as a company claim instead. Conversely, the market may be carrying information this record lacks: the September 19 spike to 74¢ and the sibling at 71¢ for July 2027 suggest some traders have views on announcement timing not visible in public reporting. The market's own path is informative — 68¢ on September 13 falling to 40¢ today, through 74¢ on September 19, describes a market that spiked on the Hodge rumor and has since drifted down as no paper materialized, which is movement in the same direction as the rules-based read rather than against it. Liquidity is moderate with about 2,484 open interest and only about 250 contracts traded in 24 hours, so this is weak evidence either way and does not strongly suggest the chain has missed something.
What would settle it: Whether, following any second-problem announcement, two of the five named Source Agencies — the Guardian, AP, the FT, NAS, or CMI — publish language stating the problem has been resolved by reference to a specific written work, as opposed to reporting that a lab claims to have solved it. The Guardian's and AP's exact framing in the first 72 hours after any such announcement is the observable that decides this market.
Moderate market: A usable spread and real open interest.
Market data from Kalshi, fetched September 28, 2026. Shown for reference; this research takes no position. View the question on Kalshi ↗

What Would Move It Next (6)

OpenAI publicly announces a claimed complete solution to the Hodge Conjecture, accompanied by a named manuscript and a Lean formalizationundated — The Information reported employees expect it soon, with the announcement possibly delayed
toward yes · strong
Where to watch: OpenAI's research announcements page and coverage by the Guardian, AP, and the Financial Times
The Guardian or the Associated Press publishes coverage stating that a Millennium Prize Problem other than Navier-Stokes has been resolved, by reference to a specific named written work, rather than reporting a lab's claimundated — would follow within days of any second-problem announcement
toward yes · decisive
Where to watch: Guardian science section and AP wire copy; two such agencies together satisfy the market's third settlement pathway
The Clay Mathematics Institute issues a statement on a second Millennium problem using resolution language referencing a specific written work, as distinct from the "apparently been settled" phrasing it used for Navier-Stokes on September 11, 2026
toward yes · decisive
Where to watch: claymath.org news and Millennium Problems pages
A peer-reviewed mathematics or scientific journal accepts and publishes a complete AI-lab-authored proof addressing a Millennium problem as originally formulated by CMIundated — no such submission is publicly known to be pending
toward yes · decisive
Where to watch: Journal tables of contents; announcement by the publishing journal
The Navier-Stokes priority dispute produces a reported unresolved identified error, retraction, or withdrawal of the OpenAI manuscript
toward no · moderate
Where to watch: arXiv or manuscript revision history, OpenAI statements, and coverage by the Guardian or Nature
Late-December 2026 arrives with no second-problem written work in public existenceDecember 2026
toward no · strong
Where to watch: Absence of announcements on OpenAI and Anthropic research pages and of coverage in the five named Source Agencies

How It Could Resolve Unexpectedly

  • A settlement dispute over whether the Navier-Stokes result itself qualifies. CMI's September 11 statement that the problem "has apparently been settled" post-dates the September 8 announcement, and if the Issuance date falls between those, someone could argue that statement is an official recognition by reference to a specific written work. The word "another" in the question, the exclusion of problems solved before the time period, and the adjacent Kalshi contract excluding Navier-Stokes by name all point against this reading, but the record here does not state the Issuance date.
  • The two-Source-Agency pathway resolving YES on coverage that is softer than a true resolution finding. If the Guardian and AP both write that a lab has solved a problem, referencing its posted manuscript, a reasonable settlement reading could be YES even while CMI evaluation remains years from completion — meaning the market could resolve YES on a result the mathematical community has not accepted.
  • A December announcement followed by rapid contestation. The rules void any purported solution reported to contain an unresolved identified error before Expiration, and the Navier-Stokes episode showed contestation arriving within 12 hours, so a late-window announcement carries real voiding risk inside the remaining days.

What the Record Could Not Settle

  • The market's Issuance date, which determines definitively whether the Navier-Stokes result and CMI's September 11 statement fall inside or outside the qualifying period. This is the single largest unresolved factor and the record provided does not state it.
  • Whether any second-problem manuscript has actually been completed internally at either lab and is being held back for messaging reasons, as The Information's source suggested — the difference between "not yet produced" and "produced but unannounced" changes the window arithmetic substantially.
  • How the Guardian, AP, and FT would specifically frame a second AI Millennium announcement. Their Navier-Stokes framing leaned toward reporting a claim, but the record here does not establish whether any of the five named agencies has since adopted resolution language.
  • Whether any journal has a submission from an AI lab on a Millennium problem under review, and on what timeline, which would be the only route to the peer-reviewed pathway inside the window.

The Established Ground

Source facts the analysis is grounded in. The → chips after each fact link to the items above that rely on it.
F1
On September 8, 2026, OpenAI published a claimed Navier-Stokes solution produced by an internal model, with a 166-page manuscript and a Lean formalization, after roughly 88 hours using about 10,000 concurrent agents.
↳ This establishes that the capability to produce a Millennium-scale proof artifact in days now exists, which is what justifies any large upward adjustment off a historical base rate.
Verified
F2
CMI's September 11, 2026 statement said the Navier-Stokes problem "has apparently been settled" and that its evaluation process is "deliberately unhurried"; CMI rules require publication in a qualifying outlet, at least two years elapsed, and general acceptance before a problem is officially resolved.
↳ The CMI-recognition and peer-review pathways to YES are structurally unavailable inside a 94-day window, forcing the forecast onto the two-Source-Agency pathway.
Verified
F3
The Information reported on September 17, 2026, citing a single OpenAI source, that employees expect to solve the Hodge Conjecture soon, but that any announcement could be delayed while the company works out how to collaborate with the math community after the Navier-Stokes backlash.
↳ This is the only concrete second-problem signal, and it carries its own delay mechanism, cutting both directions within the same fact.
Verified
F4
As of mid-to-late September 2026 coverage, there is no named paper, no Lean artifact, and no lab confirmation for a Hodge or Birch and Swinnerton-Dyer result, and CMI continues to list both as open; the BSD chatter traced to a single X user's explicitly labeled prediction.
↳ With no written work in existence yet, the clock for any qualifying recognition event has not even started, which caps how much the reported-effort signal can lift the forecast.
Verified
F5
Nature's September 8 coverage treated the OpenAI result as a company claim rather than an adjudicated solution, and Wikipedia records the Navier-Stokes result as unverified by CMI or the independent mathematical community and subject to a priority dispute.
↳ The 2026 precedent shows that even a maximally publicized AI proof produced coverage framing it as a claim, which is precisely the framing that fails the two-Source-Agency "report that the problem has been resolved" test.
Verified
F6
OpenAI's internal model scored close to 50% on the company's own benchmark of open math problems versus about 10% for the just-released GPT-6 Astra, and the model had been in training only since late August 2026.
↳ A documented step-change of this size is the strongest argument that the historical base rate for Millennium-problem resolution understates the current period.
Verified

Facts & Figures (16)

The claims behind this analysis, each with its verification status — including what is contested, unverified, or could not be established. What each grade means
Millennium Prize Problems achieving qualifying recognition (peer-reviewed publication of a complete solution, or official CMI/governing-body statement of resolution) per unit time, across the 26 years since the problems were posed in 2000: 1 of 7 problems has ever been officially declared solved (the Poincaré conjecture, prize awarded 2010), across 26 years
This is the correct denominator for the recognition event the rules require, and it captures the multi-year institutional clock. It badly breaks on the numerator side, because it is drawn almost entirely from a pre-AI era in which no actor could generate a candidate proof in days — which is what the adjustments exist to correct.
○ COUNTED BASE RATECounted from the cases listed; CMI and Wikipedia both record Poincaré as the only officially declared solution. · cases: Poincaré conjecture — Perelman preprints 2002-2003, CMI prize awarded 2010 — is the sole officially resolved case. The other six (Riemann, P vs NP, Yang-Mills, Navier-Stokes, Hodge, Birch and Swinnerton-Dyer) had no qualifying recognition event through the start of this window.
Reported-imminent AI lab Millennium-problem results converting to a qualifying recognition event within 90 days of the report, during the September 2026 episode: 0 of 2 reported near-solutions (Hodge via The Information; BSD via social media) has produced a paper, code, or institutional confirmation as of late September 2026
Directly on-point for the actual causal question, but the sample is two open cases, so it cannot bear weight as a prior. Useful mainly as a sanity check that nothing has converted yet.
— ESTIMATED BASE RATEEstimated — the episode is too new and the sample too small (n=2, both unresolved) to count a rate. Stated as estimated rather than dressed up as a base rate. · cases: The Hodge Conjecture report (The Information, single OpenAI source, September 17) and the BSD chatter (traced to one X user's labeled prediction) are the only two instances; neither had converted to a written work as of the latest coverage.
The Navier-Stokes precedent: time from AI lab announcement to a qualifying recognition event under these rules: 0 of 1 — 20 days elapsed from the September 8 announcement to September 28 with no peer-reviewed publication and no official statement of resolution
The single most informative case, because it is the same labs, the same rules environment, and maximum media attention. It shows the announce-to-recognition gap is long even in the best case. n=1 limits it to a corroborating class rather than the prior.
○ COUNTED BASE RATECounted from the single case; CMI statement and Nature coverage in grounding. · cases: OpenAI Navier-Stokes, announced September 8, 2026: manuscript and Lean formalization public; CMI's September 11 statement said "apparently been settled" with evaluation "deliberately unhurried"; Nature framed it as a company claim; no journal publication.
Demonstrated capability to produce a Millennium-scale proof artifact in days: On September 8, 2026, OpenAI published a 166-page manuscript and Lean formalization for a Navier-Stokes finite-time singularity, produced in roughly 88 hours by about 10,000 concurrent agents on an internal model that had been in training only since late August 2026 and scored close to 50% on the company's internal open-problem benchmark versus about 10% for GPT-6 Astra.
The 26-year base rate is drawn from a period in which no actor could generate a candidate Millennium proof in days, so the reference class cannot contain this mechanism at all.
✓ DOCUMENTEDOpenAI announcement and Quanta Magazine, September 8, 2026; benchmark figures per OpenAI researcher quoted in September 2026 coverage · decisive toward yes
A named second-problem effort reported as near completion: The Information reported on September 17, 2026, citing a single source at OpenAI, that employees expect to soon crack the Hodge Conjecture; OpenAI separately confirmed to the New York Times that it had made "substantial progress" on another Millennium Prize problem without naming which.
The historical class contains no instance of an actor publicly signalling an imminent second Millennium solution; this is a specific, dated signal about this window.
○ REPORTEDThe Information, September 17, 2026; OpenAI statement to the New York Times as reported in September 2026 coverage · strong toward yes
The settlement standard excludes exactly the artifacts a lab produces first: The market's additional rules state verbatim that a preprint, announcement, blog post, conference presentation, technical report, or machine-checked formalization does NOT by itself constitute a solution, and require either peer-reviewed publication, an official recognition-body statement by reference to a specific written work, or two Source Agencies reporting the problem resolved by reference to a specific written work.
The base rate measures recognition events and so partly embeds this, but the reference class cannot capture that this specific market's settlement path is narrower than CMI's own — it excludes the Lean formalization that is the labs' main credibility instrument.
✓ DOCUMENTEDMARKET RECORD, additional rules, verbatim · strong toward no
The 20-day Navier-Stokes record shows announcement does not convert to recognition on this timescale: Twenty days after the September 8 announcement, there is no peer-reviewed publication; CMI's September 11 statement said only that the problem "has apparently been settled" and described its evaluation as "deliberately unhurried"; Nature treated the result as a company claim; Wikipedia records it as unverified by CMI or the independent mathematical community and subject to a priority dispute.
This is a new, same-regime observation post-dating the reference class entirely, and it is specifically about conversion speed under maximum attention rather than about base incidence.
✓ DOCUMENTEDCMI statement, September 11, 2026; Nature, September 8, 2026; Wikipedia Clay Mathematics Institute entry as of late September 2026 · moderate toward no
No written work exists yet for any second problem: Coverage through late September 2026 records no named paper, no code, and no statement from any authoritative institution on the Hodge Conjecture; CMI continues to list the remaining problems as open; the Birch and Swinnerton-Dyer claim traced back to a single X user's explicitly stated prediction rather than to any announcement from Anthropic or OpenAI.
The reported-effort adjustment above captures the intent signal; this captures the separate fact that the clock on any recognition pathway has not started, which is independent of whether the effort succeeds.
✓ DOCUMENTEDFact-checking coverage and Clay Mathematics Institute problem listings, September 2026 · moderate toward no
Reported intent to delay the announcement for messaging reasons: The Information's source said it could take the company longer to announce a solution because it is trying to figure out how to collaborate with the math community to make the announcement without triggering another public-relations problem, following the September backlash in which several mathematicians characterized the Navier-Stokes episode as research misconduct.
A deliberate announcement delay is a mechanism specific to this window and this actor, created by the September backlash; no historical case involved a solver with a reputational reason to hold back.
○ REPORTEDThe Information, September 17, 2026, as paraphrased in Gizmodo and Decoder coverage, September 2026 · slight toward no
A second lab independently active on the remaining problems: Anthropic reported on September 4, 2026 that Claude had formalized an existing proof of Fermat's Last Theorem using about 29,500 intermediate theorems, and September 2026 commentary from someone describing knowledge of mathematicians at both labs reported that a substantial fraction of 2026 mathematical progress is known only to OpenAI and Anthropic insiders.
The base rate implicitly assumes one-solver-at-a-time academic effort; two well-resourced labs racing raises the arrival rate independently of either one's specific reported progress.
○ REPORTEDSeptember 2026 coverage of Anthropic's Fermat formalization; LessWrong commentary, September 13, 2026 · slight toward yes
On September 8, 2026, OpenAI published a claimed Navier-Stokes solution produced by an internal model, with a 166-page manuscript and a Lean formalization, after roughly 88 hours using about 10,000 concurrent agents.
This establishes that the capability to produce a Millennium-scale proof artifact in days now exists, which is what justifies any large upward adjustment off a historical base rate.
CMI's September 11, 2026 statement said the Navier-Stokes problem "has apparently been settled" and that its evaluation process is "deliberately unhurried"; CMI rules require publication in a qualifying outlet, at least two years elapsed, and general acceptance before a problem is officially resolved.
The CMI-recognition and peer-review pathways to YES are structurally unavailable inside a 94-day window, forcing the forecast onto the two-Source-Agency pathway.
The Information reported on September 17, 2026, citing a single OpenAI source, that employees expect to solve the Hodge Conjecture soon, but that any announcement could be delayed while the company works out how to collaborate with the math community after the Navier-Stokes backlash.
This is the only concrete second-problem signal, and it carries its own delay mechanism, cutting both directions within the same fact.
As of mid-to-late September 2026 coverage, there is no named paper, no Lean artifact, and no lab confirmation for a Hodge or Birch and Swinnerton-Dyer result, and CMI continues to list both as open; the BSD chatter traced to a single X user's explicitly labeled prediction.
With no written work in existence yet, the clock for any qualifying recognition event has not even started, which caps how much the reported-effort signal can lift the forecast.
Nature's September 8 coverage treated the OpenAI result as a company claim rather than an adjudicated solution, and Wikipedia records the Navier-Stokes result as unverified by CMI or the independent mathematical community and subject to a priority dispute.
The 2026 precedent shows that even a maximally publicized AI proof produced coverage framing it as a claim, which is precisely the framing that fails the two-Source-Agency "report that the problem has been resolved" test.
OpenAI's internal model scored close to 50% on the company's own benchmark of open math problems versus about 10% for the just-released GPT-6 Astra, and the model had been in training only since late August 2026.
A documented step-change of this size is the strongest argument that the historical base rate for Millennium-problem resolution understates the current period.

Sources (40)

More general research
Grounded in 40 web sources · 16 facts on the ledger · 10 verified or grounded · 6 partial or attributed · how the grades work
Analysis generated by WorldbyFlow from publicly available information. WorldbyFlow does not verify claims or endorse conclusions. New here? The two-minute overview.