Scan: Will AI solve 3 more Millennium Prize Problems in 2026?

1. Planner Fable 5.1 · Fermi decomposition + tool plan

Tools chosen: polymarket_direct polymarket_related kalshi_related article_search claude_news gdelt_news wikipedia code_execution

Sub-questions (Fermi decomposition)

What is the current Polymarket YES price for the 3-problem threshold, and how do the 1, 2, 4, 5 thresholds in the same event price out (implied distribution)?
What happened with Navier-Stokes (why is it excluded) — did an eligible lab (e.g., Google DeepMind) announce a solution, and how was that announcement worded relative to this market's 'solved' standard?
Have any eligible AI labs (Google DeepMind, OpenAI, Harmonic, Axiom Math, Math Inc, etc.) claimed or hinted at progress on Riemann, P vs NP, BSD, Hodge, or Yang-Mills in the past 6 months?
What is the current frontier of AI in research-level mathematics (Erdős problems, IMO gold, Lean formalization) and how far is it from Millennium-level open problems according to expert mathematicians?
How aggressive are eligible labs in making unverified 'solved' announcements, and is there precedent for a lab claiming a full solution to a famous open problem that was later disputed?
What is the base rate: how many Millennium problems have been resolved since 2000 (Poincaré only), and what would 3 announcements in one year imply?

Planner reasoning

This is a sibling market in a multi-threshold Polymarket event; the 3-problem threshold is extreme relative to the base rate (zero Millennium problems solved by anyone in 25 years, and the resolution requires 3 distinct problems in one calendar year). The key uncertainty is not mathematical truth but whether labs (especially aggressive ones like Harmonic, Axiom Math, Math Inc, or Google DeepMind) make bold 'solved' announcements, since no Clay verification is required. Research should anchor on the current price, check the Navier-Stokes context that prompted the market, and survey the actual state of AI-driven progress on the five qualifying problems.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.4s	1	## This Market's Polymarket Data Will AI solve 3 more Millennium Prize Problems in 2026? - Current price (probability): 7.15% - 7-day price change: -4.85% - 30-day price change: -4.85% - Total volume: $16,245 (USD notional) - Price range: 6.95% - 12.00% - Data points: 6 days
polymarket_related	OK	2.2s	0	Scanned 100 active Polymarket markets, kept 0 matches. keyword 'Millennium Prize': 0 markets \| keyword 'Riemann Hypothesis': 0 markets \| keyword 'P vs NP': 0 markets \| keyword 'Navier-Stokes': 0 markets \| keyword 'AI solve math': 0 markets
kalshi_related	OK	2.1s	1	1 related markets / summaries. keyword 'Millennium Prize': ok \| keyword 'Riemann Hypothesis': no matches \| keyword 'AI math': no matches
article_search	OK	46.8s	21	Fetched 21 unique articles across 5 queries (lookback=180d). 'AI Millennium Prize Problem solved announcement': 10 articles \| 'DeepMind Navier-Stokes AI solution': 10 articles \| 'Riemann Hypothesis AI proof lab': 10 articles \| 'Harmonic Axiom Math Inc AI mathematics breakthrough': 10 articles \| 'AI
claude_news	OK	32.7s	14	Based on my research, here are the key findings on AI progress toward Millennium Prize Problems as of mid-September 2026: - Google DeepMind's earlier fluid dynamics work (2025-2026): DeepMind, with NYU, Stanford, Brown, and mathematician Javier Gómez Serrano, used physics-informed neural networ
gdelt_news	OK	178.4s	10	GDELT: 10 articles across 3 queries (lookback=120d). 'Millennium Prize Problem AI': error HTTPSConnectionPool(host='api.gdeltproject.org', port=443): Max retries exceeded with url: /api/v2/doc/doc?query=%28millennium+OR+prize+OR+problem%29+sourcelang%3Aenglish&mode=ArtList&format=json&startdatetime=
wikipedia	OK	0.1s	5	Fetched 5 Wikipedia entries (1 missing pages).
code_execution	OK	18.4s	0	## Key Findings P(at least 3 of 5 independent problems announced "solved") by per-problem probability p: \| p \| P(≥3 of 5) \| \|---\|---\| \| 0.01 \| 0.0010% \| \| 0.02 \| 0.0078% \| \| 0.05 \| 0.1158% \| \| 0.10 \| 0.8560% \| \| 0.20 \| 5.7920% \| \| 0.30 \| 16.3080% \| - Even at an aggressive per-problem probabil

3. Evidence Brief Sonnet 5 · 7141 chars

# Current state As of mid-September 2026, OpenAI has made a disputed, unverified claim of solving Navier-Stokes (explicitly excluded from this market's qualifying list). No eligible lab has announced a full "solved" claim on any of the five qualifying problems (BSD, Hodge, P vs NP, Riemann, Yang-Mills); Anthropic's Riemann work is an explicit partial-bound improvement, not a solution. The market resolves YES only if 3 distinct qualifying problems get solved-announcements by Dec 31, 2026. # Timeline of key events - 2026-03-26 (confirmed): Google publishes efficiency research ("TurboQuant"), unrelated to Millennium problems. - 2026-08-10/13 (confirmed): Anthropic reports Claude improved a lower bound on Riemann zeta zeros from 41.6% to 67.2% — explicitly not a proof (anthropic.com, theaiinsider.tech). - 2026-09-08 (confirmed): OpenAI announces a 10,000-agent AI swarm found a Navier-Stokes singularity/counterexample, Lean-formalized; explicitly does not claim the Clay Prize (Quanta, CNBC, BBC, Guardian). - 2026-09-08/13 (confirmed): Priority dispute erupts — mathematicians Buckmaster and Alpöge (Anthropic-linked) allege OpenAI's work drew on their unpublished/rumored results (BBC, WaPo, NY Post, ABC Australia). - 2026-09-09 (confirmed): Wikipedia/Clay Institute status: Navier-Stokes remains "active," unverified, subject to dispute — not officially confirmed solved by Clay. - 2026-09-11 (reported): NYT reports OpenAI claims "substantial progress" on a second, unnamed Millennium problem (rumored Hodge or BSD) — no confirmation of which problem or that it's solved. - 2026-09 (ongoing): No qualifying-problem (BSD, Hodge, P vs NP, Riemann, Yang-Mills) solution announcement exists as of this research snapshot. # Event Will an eligible AI lab announce solutions to 3 or more of the 5 qualifying Millennium Prize Problems (BSD, Hodge, P vs NP, Riemann, Yang-Mills; Navier-Stokes excluded) by Dec 31, 2026? # Outcomes to forecast Yes / No # Kalshi market anchor No direct Kalshi-ticker quote was returned by tools for this event. The closest available live pricing is Polymarket's mirrored market on the identical question: **YES = 7.15%**, down from a 30-day high of 12.00% (7-day and 30-day trend both -4.85pp), on thin volume (~$16.2K total, 6 data points). Treat this as the best available consensus proxy; Kalshi's own order book was not retrieved. # Sub-question answers 1. **Polymarket thresholds** — Only the "3+" market was retrieved directly (7.15% YES); no data on 1/2/4/5-threshold buckets was returned, so the implied distribution across thresholds is unknown from this research. 2. **Navier-Stokes exclusion** — OpenAI (Sept 8, 2026) claimed a Lean-formalized counterexample/blowup singularity for Navier-Stokes, explicitly declining to claim the Clay Prize; Clay still lists the problem "active" and it's mired in an unresolved priority dispute with Buckmaster/Alpöge (Wikipedia, Quanta, CNBC). It's excluded from this market presumably because it's the one problem already contested/near-resolved, making it a poor test of "3 more." 3. **Recent claims/hints on qualifying problems** — Anthropic (Aug 2026): partial bound improvement on Riemann only, not a solution. OpenAI (Sept 11, 2026, per NYT): unconfirmed "substantial progress" on a second unnamed problem (rumored Hodge/BSD). No claims found on P vs NP or Yang-Mills. BSD/Hodge described as having "the least visible reinvigoration" (securemachinery.com). 4. **Frontier of AI math** — AI now solves 7/10 curated hard open problems near-flawlessly at low compute cost (securemachinery.com) and improves bounds on Riemann, but experts distinguish this from full Millennium-level proofs; no mathematician source in this research asserts Millennium problems are near AI-solvable outside Navier-Stokes. 5. **Precedent for disputed 'solved' claims** — Yes: OpenAI's Navier-Stokes announcement itself is the precedent — immediately disputed by rival researchers (Buckmaster, Alpöge/Anthropic) over originality/credit, unverified by Clay as of Sept 2026, showing labs are willing to announce aggressively before independent confirmation. 6. **Base rate** — Only 1 of 7 Millennium problems (Poincaré, 2003/2010) solved in ~25 years (~4%/year for any one; ~0.57%/problem-year). Code-execution model: even at an aggressive 30% per-problem annual solve probability, P(≥3 of 5) ≈16%; at realistic ≤10%, <1%; using historical base rate, ~1.85×10⁻⁶. # Key facts (high-confidence, factual) 1. [Wikipedia] Only Poincaré Conjecture officially solved since 2000. 2. [Quanta/CNBC/BBC] OpenAI's Navier-Stokes claim (Sept 8, 2026) is unverified by Clay and excluded from this market. 3. [Anthropic] Claude's Riemann result is a bound improvement, not a proof. 4. [securemachinery.com] No 2026 AI breakthroughs reported on BSD, Hodge, P vs NP, or Yang-Mills. 5. [Polymarket] Mirror market YES price fell from 12% to 7.15% over 30 days. # Cross-market signals - Kalshi related: "Which Millennium Prize Problem will be solved next — BSD" priced at 92% YES (KXMILLENNIUMNEXT-45-BSD) — reflects near-certainty BSD (or a specific problem) will be *first/next* solved eventually, not a 2026 timeframe signal; not directly comparable. - Polymarket: Same-question mirror at 7.15%, declining trend, low volume/liquidity — weak but real signal of skepticism. - Sportsbook implied: None found. # Analyst opinions and speculation - Zvi (thezvi.wordpress.com): notes prediction markets gave AI <40% chance of solving even ONE Millennium problem before 2030 as recently as April 2026 — 3 in 2026 seen as far-fetched. - Manifold market on "AI solves a Millennium Prize Problem in 2026" (single problem) traded ~38% pre-OpenAI announcement. - Commentators split: some see Navier-Stokes claim as proof AI is a serious math-research force; others (Buckmaster) call the "race to solve" framing "pointless" and dispute originality. # Directional lean per outcome - **Yes**: OpenAI's rapid, aggressive announcement cadence (Navier-Stokes + rumored second problem) and demonstrated 10,000-agent scaling show labs can move fast; Anthropic and OpenAI both actively working adjacent problems. - **No** (dominant): Zero qualifying-problem full solutions exist yet; excluded Navier-Stokes claim is itself disputed/unverified; historical base rate is ~1 solved problem per 25 years; quantitative modeling puts realistic P(≥3) at well under 5%; Polymarket pricing (7.15%, declining) concurs. # Gaps / unknowns - No direct Kalshi order-book data for this specific ticker was retrieved. - Polymarket thresholds for 1/2/4/5 problems not obtained — can't confirm implied full distribution. - Unclear if/which second problem OpenAI is allegedly working on; unconfirmed rumor only. - No mathematician expert quotes directly assessing near-term feasibility of Riemann/P vs NP/Hodge/BSD/Yang-Mills solutions. # Calibration anchors - Polymarket mirror YES price: 7.15% (declining from 12% high) — primary available anchor. - Historical base rate: 1 Millennium solution in 25 years; quantitative model puts P(≥3 in one year) at <2% under any realistic assumption, near-zero under historical base rate.

4. Forecasters Fable 5.1 + GPT-5.6 Sol · in parallel

claude-fable-5-1

72% conf

Yes	4%
No	96%

As of mid-September 2026, zero of the five qualifying problems (BSD, Hodge, P vs NP, Riemann, Yang-Mills) have any full solution announcement; Anthropic's Riemann work is an explicit bound improvement and OpenAI's rumored second-problem 'substantial progress' is unconfirmed and not a solved claim. The only precedent, OpenAI's Navier-Stokes claim, is excluded from this market and is itself disputed and unverified by Clay. Getting three distinct qualifying solved-announcements in the remaining ~3.5 months would require an unprecedented cadence far beyond the historical base rate (1 solution in 25 years) and even aggressive per-problem models put P(≥3) well under 5%. The Polymarket mirror at 7.15% (declining, thin volume) is the anchor; I shade slightly below it for longshot bias and the very short remaining window, while leaving room for aggressive, possibly disputed lab announcements.

gpt-5.6-sol

72% conf

Yes	3%
No	97%

The 7.15% Polymarket YES price is the best available consensus anchor, though its low volume and lack of direct Kalshi data limit its reliability. Navier-Stokes is excluded and remains disputed, while Anthropic's Riemann result is only a partial bound improvement and no qualifying problem has a full solution announcement as of mid-September. OpenAI's rumored progress on one unnamed qualifying problem may raise the chance of one announcement, but it provides little support for the three distinct announcements required within roughly three and a half months. Given the extremely low historical base rate, the absence of visible breakthroughs on the other problems, and the declining market price, I move YES below the 7.15% anchor, while retaining a small tail for correlated AI advances and aggressive lab announcements.

5. Devil's Advocate Sonnet 5 · red-team critique

Both forecasts shade below the 7.15% Polymarket anchor citing "longshot bias," but neither justifies why a thin, low-volume ($16.2K, 6 data points) market should be trusted enough to anchor on at all rather than defaulting more heavily to the base-rate/model calculations that put P(≥3) at <2% or even ~10⁻⁶ — the two anchoring choices (trust market vs. trust base rate) pull in different directions and neither forecast reconciles this tension explicitly.

Neither forecast grapples with the "disputed announcement" precedent as a resolution-criteria risk: OpenAI's Navier-Stokes claim shows labs will announce aggressively pre-verification, meaning if 3 disputed/unverified claims emerge on qualifying problems, the market could arguably resolve YES on announcement alone (per event framing "announce solutions"), which is a fatter tail than the base-rate math assumes.

Both forecasts treat the OpenAI "second unnamed problem" rumor as weak evidence, but neither addresses the compounding effect if OpenAI's swarm approach is validated as a generalizable method — the "7/10 curated hard open problems" solved near-flawlessly fact suggests capability growth could be nonlinear, a scenario underweighted by static historical base rates in both.

Neither forecaster flags the inconsistency between the Kalshi cross-market signal (92% YES on "which problem solved next: BSD") and their own near-zero probability for 3 solved by year-end — if BSD is seen as 92% likely to be *the next* solved (eventually), this doesn't calibrate against 2026 timing but neither forecast explains why this doesn't even marginally raise probability for at least one qualifying solve.

Both converge tightly (0.03–0.04) with identical 0.72 confidence, suggesting anchoring on each other or the same Polymarket number rather than independently stress-testing the tail scenario where AI-agent swarms produce multiple simultaneous claims in Q4 2026 (a real possibility given the Sept 8 precedent occurred just before the snapshot).

Neither addresses the ambiguity in "3 more" — since Navier-Stokes is excluded despite being AI's actual breakthrough, both forecasts implicitly assume the exclusion has no bearing on lab incentives, but a rational lab (OpenAI) knowing Navier-Stokes doesn't count toward this market has extra incentive to redirect swarm resources at the 5 qualifying problems, which is a strategic response not modeled.

6. Aggregator Fable 5.1 · final calibrated forecast

4% (72% confidence)

Yes: 4% No: 96%

With ~3.5 months left, zero of the five qualifying problems (BSD, Hodge, P vs NP, Riemann, Yang-Mills) has a full solved-announcement; Anthropic's Riemann work is an explicit partial bound and OpenAI's rumored second problem is unconfirmed 'substantial progress,' not a solution. The critique correctly notes that announcement-based resolution and aggressive lab behavior (the disputed Navier-Stokes claim) create a fatter tail than pure historical base rates imply, but even generously assuming OpenAI announces one qualifying claim in Q4, reaching three distinct problems requires two further independent breakthroughs on problems with no visible progress. The thin Polymarket anchor (7.15%, declining) likely carries longshot bias, so shading below it toward the model-based estimates is justified while leaving room for correlated swarm-driven announcements. Final call: ~3.5% YES, in line with both forecasters.

Pipeline Timing

Total pipeline time: 281.7s

Per-tool research timings shown in the Research section above.