forecasting
agent
← all forecasts

Will P vs NP be resolved before 2030?

kalshiScience and Technology2026-11-20llm cost $1.66view on kalshi ↗
36%
Agent
53%
Market Price
-17.0pp
Edge
38%
Confidence
6,808
Volume
31.0c
Spread
1143
Days to resolution
final rationale
average of 2 draws (32%, 40%), redrawn after an extra research pass
already resolved?
The condition is not met. Clay's P vs NP page, fetched 2026-11-20, still says "Unsolved". No lab or credible researcher has claimed a full resolution; the fall-2026 ETH refutations by OpenAI and Anthropic are explicitly short of P vs NP, and Brockman himself said P probably does not equal NP.
time left
About 3 years and 1 month remain before the Jan 1, 2030 deadline.
if nothing changes
If nothing changes, the market resolves No.
path to no
AI-lab results keep landing on adjacent questions (subexponential SAT, ETH and SETH variants), but polynomial-time SAT is never found and P≠NP stays blocked by the known relativization, natural-proofs and algebrization barriers. Any premature claim gets a publicly identified error, which triggers the contract's carve-outs.
path to yes
A lab publishes a Lean-verified polynomial SAT algorithm or a P≠NP proof, two Source Agencies report it as resolved, and no flaw is found before expiry. The contract's loose third route makes this possible well before any Clay recognition. The Kalshi anchor is a 53% mid, but the book is thin and wide (38/69 bid-ask, about 445 contracts a day), and the last daily close was 23%. That makes the mid a noisy signal rather than a strong consensus. The environment does justify far more than the historical base rate (Metaculus medians of 2054–2077): two labs independently refuted ETH with Lean proofs, there is a Navier–Stokes counterexample claim, the sibling Riemann market trades at 86%, and Aaronson reports labs racing on P vs NP. Even so, P vs NP is widely seen as the hardest of these problems, so I settle near 0.32, between the market's last close and its mid.
earlier forecasts on this market
datepredictedmarket priceconfidence
2026-10-28 18% 32% 44%
1. Planner Opus 5.5
Tools chosen: kalshi_direct resolution_source kalshi_related polymarket_related wikipedia
Planner reasoning
P vs NP being resolved before 2030 is very unlikely, so the Kalshi price is the main anchor. The Clay Millennium Prize page is the resolution source. Related markets on other millennium problems or on P vs NP itself give cross-venue checks. Wikipedia covers the background and status of the problem, and an arXiv count shows how many claimed-proof papers appear (they are noise, but tell us how often claims surface).
2. Research 4 web research agents, Sonnet 5.5
summary
RESEARCH DIGEST (built in code: tool summaries and each report's bottom line; the forecasters read the full tool outputs and reports) Tools - kalshi_direct: ## This Market's Kalshi Data **Before 2030** (KXPNP-32-30JAN01) - Current price (probability): 53.00% (live mid; bid 38% / ask 69%; last trade 69%) - Last daily close (2026-11-11): 23.00% - 7-day price change: +23.00% - 30-day price change: +41.00% - Average daily volume: 445 contracts - Price range: 12.00% - 57.00% - Data points: 13 days - resolution_source: already_resolved=unclear; matches_criteria=True; last_updated=unknown (page footer shows © 2026). The Clay Mathematics Institute's P vs NP page currently labels the problem "Unsolved". On the Millennium Problems index, P vs NP appears under "Unsolved problems". The only problem listed under "Solved problems" is the Poincaré Conjecture, and Navier-Stokes is listed separately under "Active problems" - kalshi_related: 2 related markets / summaries. series KXPNP: 0 markets (skipped 6 no-signal) | keyword 'millennium prize': no matches | keyword 'Riemann hypothesis': ok | keyword 'conjecture proof': ok - polymarket_related: Polymarket search over 3 keyword(s), kept 10 matches. keyword 'P vs NP': 0 markets | keyword 'millennium prize': 5 markets | keyword 'Riemann': 5 markets - wikipedia: Fetched 2 Wikipedia entries (0 missing pages). Research loops, bottom lines - current state: As of 2026-11-20, no accepted resolution of P vs NP exists, and Clay still lists it as unsolved. AI labs are producing a stream of results in complexity theory, including the contested ETH disproof, but none claims to resolve P vs NP and none is verified. - base rate: The historical base rate is very low. P vs NP has been open for more than 50 years, none of the other Millennium problems except Poincaré has been resolved in 26 years, and expert and forecaster medians (Metaculus ~2054–2077) are decades past 2030. The residual probability comes from the Kalshi criteria being looser than Clay's and from AI-assisted proof claims. - yes case: The strongest YES case is that two frontier labs refuted ETH in the same week, with Lean proofs and fast expert uptake. Labs are reportedly targeting P vs NP, and the contract's looser media-report route could be met by a single verified AI manuscript before 2030. The case still rests on a result that does not yet exist, since ETH's refutation is explicitly not P vs NP. - no case: The NO case rests on four points. The only strong AI-lab result, on ETH, is unresolved and does not decide P vs NP. Every past high-profile claim collapsed within days to weeks. The criteria's retraction and "identified error" clauses penalise premature claims. No scheduled mechanism delivers a resolution before 2030. Crux follow-up - The evidence favours No for now. No credible full resolution exists, and the fall-2026 AI results are explicitly short of P vs NP. The residual risk is a lab-produced, Lean-checked polynomial SAT algorithm or proof appearing before 2030, which is plausible given the pace of AI results but currently unevidenced. The contract's loose third route lowers the bar for a claim to count.
tools
toolstatustimeitemssummary
kalshi_direct OK 2.0s 1 ## This Market's Kalshi Data **Before 2030** (KXPNP-32-30JAN01) - Current price (probability): 53.00% (live mid; bid 38% / ask 69%; last trade 69%) - Last daily close (2026-11-11): 23.00% - 7-day price change: +23.00% - 30-day price change: +41.00% - Average daily volume: 445 contracts - Price rang
resolution_source OK 5.6s 1 already_resolved=unclear; matches_criteria=True; last_updated=unknown (page footer shows © 2026). The Clay Mathematics Institute's P vs NP page currently labels the problem "Unsolved". On the Millennium Problems index, P vs NP appears under "Unsolved problems". The only problem listed under "Solved
kalshi_related OK 17.0s 2 2 related markets / summaries. series KXPNP: 0 markets (skipped 6 no-signal) | keyword 'millennium prize': no matches | keyword 'Riemann hypothesis': ok | keyword 'conjecture proof': ok
polymarket_related OK 0.9s 10 Polymarket search over 3 keyword(s), kept 10 matches. keyword 'P vs NP': 0 markets | keyword 'millennium prize': 5 markets | keyword 'Riemann': 5 markets
wikipedia OK 0.8s 2 Fetched 2 Wikipedia entries (0 missing pages).
web research agents 4/4 ok, 18 searches, 4 page reads
dimensionstatustimesearchespage readsreport
current state OK 19s 4 0 4301 chars
base rate OK 32s 5 1 5258 chars
yes case OK 24s 5 1 5815 chars
no case OK 26s 4 2 6246 chars
report: current state
- [2026-11-20] Clay Mathematics Institute's Millennium Problems page (fetched) still lists P vs NP under "Unsolved problems". Navier-Stokes is the only problem listed as "Active", and Poincaré is the only one listed as solved. https://www.claymath.org/millennium-problems/ - [2026-11-17] Emanuele Viola (Northeastern) wrote in CACM that the fall-2026 AI results "did not prove that P equals NP". Per his post, OpenAI released a large batch of results on Oct 6 and a claimed disproof of the Exponential Time Hypothesis (ETH) on Oct 26. ETH is a strengthening of P≠NP. Viola notes that "virtually all researchers" still believe P≠NP, and he personally believes P=NP. https://emanueleviola.wordpress.com/2026/11/17/what-happened-in-fall-2026-ai-and-p-vs-np/ - [2026-10-28] Commentary says OpenAI's ETH claim does not prove P=NP, because ETH can be false while P≠NP. The same piece says Clay still lists P vs NP as unsolved. The ETH manuscript and a Lean formalization are published, and independent examination was still under way. https://www.ai.joaoqueiros.com/blog/openai-eth-proof-wes-roth-p-vs-np-what-it-means - [2026-10-27] A report says the ETH claim has not been verified. It also says Anthropic claimed on Oct 16 that an internal model found a core 3-SAT algorithm, which it calls an unverified company claim. Per this report, a Manifold market on "Does P = NP" was at about 12% after the news. That is a different venue and a different question, and the page was blocked, so I could not fetch it. https://tech-insider.org/does-p-np-manifold-market-odds-volume-spike-2026/ - [2026-10-26] The ETH-falsity manuscript is dated Oct 25, 2026 (OpenAI), as cited by Viola. https://emanueleviola.wordpress.com/2026/11/17/what-happened-in-fall-2026-ai-and-p-vs-np/ - [2026-10-03] Cairn Commons lists P vs NP as open, with no accepted proof either way. https://cairn-commons.com/problems/p-vs-np - [2026-09-08] OpenAI claimed an AI-generated Navier-Stokes counterexample, with a Lean proof, and said it would not claim the prize. Wikipedia says the result has not been independently verified. Clay's rules require publication in a qualifying outlet, general acceptance, and at least two years since publication before it considers a solution. https://en.wikipedia.org/wiki/Millennium_Prize_Problems - [2026-09-11] MathWorld says Clay lists five problems as unsolved and labels Navier-Stokes "Active". It also restates the two-year rule. https://mathworld.wolfram.com/MillenniumPrizeProblems.html - [2026-08-07] Proof Atlas lists P vs NP as "Recent proof claim under review". The claims are an arXiv preprint (2606.03194, June 2) claiming a Lean-verified P=NP, and a July package claiming P≠NP. Both rely on axioms. Proof Atlas found no accepted machine-checked resolution. https://www.proofatlas.ai/collaboration/p-vs-np/ - [2026-07-19] A Zenodo survey, "P versus NP in 2026", says no proof of either direction is claimed and describes the field's barriers. The survey is not externally peer reviewed. https://zenodo.org/records/21440860 - [2026-11-14] Britannica's P vs NP entry was updated and still describes the problem as unresolved. https://www.britannica.com/science/P-versus-NP-problem - [2026-02-02] A Zenodo upload ("Interface Geometry") claims P=NP is "resolved". It is a self-published, unreviewed crank-type claim, and no source treats it as credible. https://zenodo.org/records/18464652 - Scheduled before close: no Clay decision or journal publication on P vs NP is scheduled. Clay's two-year waiting rule means a Clay prize cannot come before 2028 even for a paper published now. The question's own criteria use other routes, such as peer-reviewed publication, recognition by an authoritative body, or two Source Agency reports. BOTTOM LINE: As of 2026-11-20, no accepted resolution of P vs NP exists, and Clay still lists it as unsolved. AI labs are producing a stream of results in complexity theory, including the contested ETH disproof, but none claims to resolve P vs NP and none is verified. NOT FOUND: Any claim by OpenAI, Anthropic or another AI lab to have resolved P vs NP itself. Any statement from Clay, the AMS or the IMU on P vs NP. Any scheduled event that would settle the question before 2030. Independent verification of the ETH disproof or of Anthropic's 3-SAT claim.
report: base rate
- [2026-11-20] Clay's rules say it will not consider a proposed solution until three conditions are met. The solution must be published in a Qualifying Outlet, at least two years must have passed since publication, and it must have received general acceptance in the global mathematics community. This matters for the Kalshi contract. Kalshi's own criteria are looser and don't require Clay's award: one peer-reviewed publication, one authoritative-body statement, or two Source Agency reports can each suffice. https://www.claymath.org/millennium-problems/rules/ - [2026-09] The Clay Mathematics Institute says only the Poincaré Conjecture, of its seven problems, has been resolved. This comes from the Clay lecture-series page, which covers 17 Sep 2025 to 15 Apr 2026. https://www.claymath.org/events/millennium-prize-problems-lecture-series/ - [2026] Reference-class timing: Perelman's Poincaré papers were posted in 2002–2003. The Millennium Prize was awarded on 18 Mar 2010, about 7–8 years later, and Perelman declined it. One of seven problems has been resolved in the 26 years since 2000, and P vs NP, which dates from 1971, is still open after more than 50 years. https://en.wikipedia.org/wiki/Millennium_Prize_Problems - [2026-08-07] ProofAtlas, checked 7 Aug 2026, lists two recent AI/Lean claims. A June 2026 arXiv preprint (Arthanari, 2606.03194) claims P = NP. A July 2026 Lean package claims P ≠ NP. Both rely on axioms, were not independently audited, were not published in a qualifying outlet, and Clay still lists the problem as unsolved. Zenodo and GitHub amateur claims also appear. Claimed proofs are routinely posted, and none has been accepted. https://www.proofatlas.ai/collaboration/p-vs-np/ - [2026-11-17] A CACM blog post (snippet only; the page was behind a bot check and I could not fetch it) says the P vs NP problem "has not yet been solved". It reports that OpenAI claimed on Oct 26 to have disproved the Exponential Time Hypothesis, a strengthening of P≠NP. The author says the AI results did not prove P = NP. https://cacm.acm.org/blogcacm/what-happened-in-fall-2026-ai-and-p-vs-np/ - [2026-10] A MindStudio summary says OpenAI's claims are not independently verified and leave P vs NP untouched. It says Greg Brockman noted P probably does not equal NP. This is a secondary source. https://www.mindstudio.ai/blog/openai-ai-discoveries-math-physics-cs - [2018] The third Gasarch poll had 124 responses, and 88% of respondents believed P ≠ NP (61% in 2001, 83% in 2011). It also asked when the problem would be resolved. The 2002 poll's individual answers included 2100 and 2525, and one 2018-poll respondent kept the original poll's 2036 guess. Experts overwhelmingly expect a very long timeline. https://en.wikipedia.org/wiki/P_versus_NP_problem and https://www.cs.umd.edu/users/gasarch/BLOGPAPERS/pollpaper3.pdf - [2026] Metaculus's community forecast (snippet, undated) for "When will P = NP be settled?" gives a median of Jan 2074, with an interval starting about Sep 2041. A separate Metaculus question on the six remaining Millennium problems gives a median of about Aug 2054 for P vs NP. Another Metaculus page shows 2077. Metaculus puts essentially no median mass before 2030. https://www.metaculus.com/questions/9448/date-p-np-is-settled/ , https://www.metaculus.com/questions/15838/millennium-prize-problems-solution-date/ , https://www.metaculus.com/questions/1408/p--np-is-true/ - [2026] For comparison, Metaculus puts the Riemann Hypothesis at about Jan 2033 on the Millennium question (snippet). Metaculus gives Collatz a median of 2055 and Riemann a median of April 2044 on separate questions. https://www.metaculus.com/questions/15838/millennium-prize-problems-solution-date/ , https://www.metaculus.com/questions/7501/date-twin-prime-conjecture-is-resolved/ - [2026-10] Kalshi lists sibling "Before 2030" markets for the Riemann hypothesis winning the Millennium Prize and for the next Millennium Prize being awarded. The page is undated and shows prices I did not use. https://kalshi.com/category/science - [2026-11-20] The Kalshi criteria's third route needs only two "Source Agency" reports of a proof, with no uncorrected identified error. That is a lower bar than Clay acceptance. A widely reported but wrong claim is the main route to a Yes resolution that isn't a true resolution. Even so, no claim has met it historically, and several have been reported and then refuted (for example Deolalikar in 2010). The Deolalikar example is from my background knowledge, not from a fetched source. https://marketline.app/market/KXPNP-32-30JAN01 BOTTOM LINE: The historical base rate is very low. P vs NP has been open for more than 50 years, none of the other Millennium problems except Poincaré has been resolved in 26 years, and expert and forecaster medians (Metaculus ~2054–2077) are decades past 2030. The residual probability comes from the Kalshi criteria being looser than Clay's and from AI-assisted proof claims. NOT FOUND: the Kalshi contract's actual rules text and its official Source Agency list; the dated Metaculus interval and number of forecasters (page blocked); the CACM article text; the Gasarch 2018 poll's median resolution date as a number; any Polymarket or Manifold P vs NP market.
report: yes case
YES means: Between now and the Jan 2030 close, P vs NP (in the Clay formulation) is resolved by a peer-reviewed proof or disproof, by a written statement from an authoritative body, or by two Source Agency reports of a proof with no uncorrected error. It also requires that the resolution is not retracted before expiry. - [2026-10-28] Lance Fortnow wrote that it is "safe to assume" ETH is now "toast". He had not verified the proof himself. Ryan O'Donnell and Shiva Kintali posted simplified follow-up proofs within days. Impagliazzo, Paturi and Williams opened a discussion site on Oct 29. Mechanism for YES: AI-found, human-checkable complexity results are being accepted quickly by experts. The bound is 2^{O(n log log n/√log n)} for k-SAT. Fortnow notes it does not extend to all of NP. https://blog.computationalcomplexity.org/2026/10/the-exponential-time-refutation.html - [2026-11-06] Anthropic's independent paper (49 pages, Lean 4, re-checked by a second kernel) reaches essentially the same bound as OpenAI's 67-page paper. Anthropic says Claude found the core idea with no human input after about 14 million tokens. Williams says parts of OpenAI's proof look routine, with "two or three alien-looking parts". No flaw had been found. Two labs plus several human researchers reproducing the result is a stronger signal than a lone claim. It shows the frontier of AI-assisted complexity theory moving toward NP-hardness itself. https://www.datacamp.com/blog/ai-exponential-time-hypothesis - [2026-10-28] Fortnow notes a corollary posted by Subercaseaux: NTIME(T) ⊆ DTIME(2^{O(T(loglogT)^3/√logT)}). This is a nontrivial simulation of nondeterminism by deterministic time. It is the first sub-brute-force result for general nondeterministic machines in this account. It is still far from polynomial. https://blog.computationalcomplexity.org/2026/10/the-exponential-time-refutation.html - [2026-11-17] Viola, a P=NP believer, argues in CACM that the fall-2026 results show experts underestimate efficient computation. He lists faster algorithms for problems conjectured not to admit them. This is the one credible dissenter's case that P=NP could be near. Note that P=NP would be the likelier "resolution" via a polynomial 3-SAT algorithm. That is a single artifact that AI search could find and Lean could check. https://cacm.acm.org/blogcacm/what-happened-in-fall-2026-ai-and-p-vs-np/ - [2026-10-27 to 2026-11-18] Scott Aaronson said on a podcast that OpenAI and Anthropic are "now racing to solve" P vs NP with internal models, but have not yet solved it. The host raised "sooner than 2029". This is a mainstream theorist describing the labs as actively targeting it. In contrast, Aaronson said on 2026-09-15 that AI progress gives no evidence for P=NP. https://moderncto.io/986-theoretical-computer-scientist-ai-just-changed-math-forever-with-scott-aaronson/ - [2026-11-18] Dan Boneh, at an a16z Frontier discussion, raised the hypothetical of AI solving P vs NP. He said a machine-checkable proof could be too long for people to follow. Relevance: the Kalshi criteria's third route needs reports of a "publicly available written work". A Lean-verified AI proof could satisfy that without human comprehension or peer review. https://www.tokenpost.com/news/technology/61107 - [2026-10-26] OpenAI's own post says the ETH work was "subsequently verified in Lean". It routes verification through an academic group. This is a template for a future P vs NP claim: Lean proof, then expert shepherding, then two or more media reports. That would be quick against the two-year Clay wait. The OpenAI page returned 403 on fetch, so this comes from search snippets. https://openai.com/index/building-scientific-understanding-with-academic-communities/ - [2026-10-06] OpenAI's Oct 6 release covered 23 Erdős problems. Terence Tao and others criticised the volume. The lab's cadence is now "weeks, not years", and the datacamp piece says new AI results land every week or two. Most successful paths to a third-route YES run through volume of claims and speed of media uptake. https://officechai.com/ai/new-openai-research-says-exponential-time-hypothesis-eth-is-false/ - [2026-10-26] The Kalshi contract's literal criteria need no Clay award. The third route needs only two Source Agency reports and no identified uncorrected error. The single most likely YES path is a lab-published, Lean-verified P≠NP or P=NP manuscript. Its prerequisites are: (1) a lab or group publishes a full proof, (2) two named Source Agencies report it as resolved, and (3) no flaw is publicly identified before expiry. P=NP via a polynomial 3-SAT algorithm is the most checkable form. A P≠NP proof is hard because of known barriers, and experts say they lack even a strategy. The Source Agency list could not be confirmed. - [2026-11-20] The contract title says "resolved", and the criteria's route 3 says "otherwise resolved". The marketline page for this contract (KXPNP) reproduces the three-route text. A lenient resolver could count a widely reported Lean-verified claim before it has been vetted. This is an interpretive risk, not a confirmed fact. https://marketline.app/market/KXPNP-32-30JAN01 BOTTOM LINE: The strongest YES case is that two frontier labs refuted ETH in the same week, with Lean proofs and fast expert uptake. Labs are reportedly targeting P vs NP, and the contract's looser media-report route could be met by a single verified AI manuscript before 2030. The case still rests on a result that does not yet exist, since ETH's refutation is explicitly not P vs NP. NOT FOUND: Any lab claim to have resolved P vs NP itself. Any outcome of the Impagliazzo-Paturi-Williams review. The contract's official Source Agency list. Any journal acceptance of the ETH papers. Any Aaronson quote that gives a date for P vs NP.
report: no case
YES means: P vs NP, in the Clay formulation, is resolved before Jan 1, 2030 by one of three routes: a peer-reviewed proof or disproof, a written statement from an authoritative body, or two Source Agency reports of a specific proof with no uncorrected error. The resolution must not be retracted before expiry. This report argues the NO side only. - [2026-11-07] The ETH disproof is the strongest AI result in this area, and it still falls short of P vs NP. Impagliazzo, one of ETH's authors, said it is "not 100% certain". The Lean proofs were replayed by outsiders, but the human verdict from Impagliazzo, Paturi and Williams was still pending. The sharper bound was not fully formalized. Outside estimates put the claimed advantage only beyond about 10^2,389 variables. A subexponential 3-SAT algorithm does not touch P vs NP. If AI's best result is a hypothesis stronger than P≠NP, a full resolution is a much larger step. https://shiftdelete.net/en/exponential-time-hypothesis - [2026-10-27] The Washington Times reported that OpenAI's own statement framed the ETH result as an open question under human verification. It said any faster algorithms would need "significant further research". https://www.washingtontimes.com/news/2026/oct/27/openai-says-model-refuted-exponential-time-hypothesis/ - [2026-10-27] A Hacker News thread on the OpenAI ETH PDF shows a pattern relevant to the "two Source Agency reports" route. Commenters first read the result as bearing on P vs NP, then corrected themselves. They found the runtime bounds do not survive polynomial reductions, so the result does not even separate NP from EXP. Early media framing overstates results, which is a risk. But the later corrections, and the lack of any lab claiming P vs NP, are what make a clean false-Yes unlikely. https://news.ycombinator.com/item?id=50251661 - [2024-01] A survey in IEEE Access counted 126 papers claiming to resolve P vs NP: 67 for P=NP, 55 for P≠NP, and 4 for independence. None was accepted. The authors argue that the review queue grows without bound as submissions arrive. Journals such as the Journal of the ACM impose a two-year resubmission ban on P/NP claimants. Formal verification is proposed as a pre-filter, which makes a quick journal route less likely. https://epub.jku.at/obvulioa/download/pdf/12261171 - [2010-08] The Deolalikar precedent shows how fast a high-profile claim fails. Cook called it "relatively serious", and it had 165,000 Google hits. The drafts were discredited and withdrawn within about ten days of the Aug 6 posting. Schneier's blog had marked it "wrong" by Sep 11. Any claim that gets two outlets to report it would likely be corrected publicly before the Jan 2030 expiry, which triggers the criteria's "identified error" and "retracted" carve-outs. https://www.cs.rice.edu/~vardi/comp409/pvnp10.pdf and https://www.schneier.com/blog/archives/2010/08/p_np_1.html - [2026-11-20] Peer-reviewed publication of a genuine proof by the end of 2029 would require a journal to complete refereeing in about 3 years, for a hard 100+ page result. The only recent lab-style claims on a Millennium problem are the Sept 8 Navier-Stokes counterexample and the Nov 20 Anthropic-mathematician Riemann claim. Neither is a P vs NP claim, and neither has been accepted. Sibling markets show the pattern: Riemann resolution is priced above Navier-Stokes and P vs NP is priced lower. https://www.bitcoin.com/prediction-markets/kalshi/when-will-p-vs-np-be-resolved--KXPNP-32/ - [2026-11-20] A different Kalshi sibling, "Which Millennium Prize Problem will be solved next?", showed Riemann Hypothesis 94% and Hodge 6%. P vs NP did not appear in that list as a likely next solution, so that market puts it behind Riemann. Same page, same snapshot date, 6 related markets. https://www.bitcoin.com/prediction-markets/kalshi/when-will-p-vs-np-be-resolved--KXPNP-32/ - [2026-01] Manifold's multi-year market "Will P vs NP be resolved by the end of these years?" listed 4% for 2030 against 31% for 2045 and 41% for a later year. That is a play-money market with 12 traders, so it is weak evidence, but it is a different venue than Kalshi. A side comment on Manifold's "all Millennium problems before 2030" market says P vs NP is "impossible to beat", which is an anecdote and not a forecast. https://manifold.markets/BoltonBailey/will-p-vs-np-be-resolved-by-the-end-ht9LR2lRSs and https://manifold.markets/SG/all-millennium-prize-problem-solved - [2026-09-10] Tech Insider's Manifold item shows that a single contested lab claim moved a "solved before 2030" market to 24%, while the Clay two-year rule alone makes the official route near-impossible on the stated timeline. It says six open problems would need accepted resolutions in under four years. https://tech-insider.org/millennium-prize-problems-solved-2030-prediction-market/ - [2026-08-07] ProofAtlas says neither of the two 2026 P vs NP artifacts was independently audited. Both leave decisive steps as axioms, so under Lean's own logic neither counts as a complete proof. Axiomatised AI "proofs" do not satisfy the criterion of a "complete proof". https://www.proofatlas.ai/collaboration/p-vs-np/ - Criteria technicality: the contract says "original form" of the Clay formulation. A result on a variant, such as a restricted model, a relativized world, ETH/SETH, or a Lean statement with different definitions, is not a resolution. The ETH case already needs a human check that its formal definitions match the textbook version. - Status quo: nothing is scheduled between now and Jan 2030 that would force a resolution. No journal issue, Clay hearing or prize committee is set for P vs NP. BOTTOM LINE: The NO case rests on four points. The only strong AI-lab result, on ETH, is unresolved and does not decide P vs NP. Every past high-profile claim collapsed within days to weeks. The criteria's retraction and "identified error" clauses penalise premature claims. No scheduled mechanism delivers a resolution before 2030. NOT FOUND: Any claim by an AI lab or mathematician to have resolved P vs NP itself. Any Clay, AMS or IMU statement on P vs NP. Kalshi's official Source Agency list. The expert verdict on the ETH disproof. The Science article text, which I did not fetch.
3. Forecasters 2 draws, Opus 5.5 + Fable 5.2, re-drawn after crux research
crux research
Crux research ran because the draws' mean 0.35 sits 0.18 from the same-question market "this kalshi market" at 0.53. One more research agent looked into that point, and every draw below was redone with its report.
first-round draws: 35%, 36% , crux research: 4 searches, 1 page read
report: crux
CRUX: Does any frontier AI lab or credible researcher currently have a complete P vs NP resolution, meaning a polynomial-time SAT algorithm or a P≠NP proof? This matters because the contract's third route, two Source Agency reports with no uncorrected error, would let a claim count without Clay's two-year wait. As of today the evidence shows only adjacent results, such as the subexponential 3-SAT algorithm. Evidence: - 2026-11-20: Clay's P vs NP page, which I fetched, still shows the status "Unsolved". https://www.claymath.org/millennium/p-vs-np/ - Clay's rules require publication in a qualifying outlet, general acceptance, and two years after publication before Clay considers a solution. Wikipedia and MathWorld both state this. This governs Clay's prize, not this contract; the contract has its own faster routes. https://www.claymath.org/millennium-problems/rules/ and https://en.wikipedia.org/wiki/Millennium_Prize_Problems - 2026-10-26: OpenAI published a preprint and a Lean formalization claiming the Exponential Time Hypothesis (ETH) is false. The algorithm is subexponential, roughly 2^(n/√log n), not polynomial. It is not peer-reviewed, and outside review, including by the ETH's authors, was under way. https://xenospectrum.com/en/openai-exponential-time-hypothesis-false-claim/ and https://beincrypto.com/openai-exponential-time-hypothesis-p-vs-np/ - 2026-10-26: OpenAI president Greg Brockman said "p still probably doesn't equal np". Reports say Anthropic also claimed an ETH refutation, a "2^o(n) Time Algorithm for 3-SAT". https://datacamp.com/blog/ai-exponential-time-hypothesis - 2026-11-17: Emanuele Viola's CACM post says the AI results "did not prove that P equals NP". He says they give faster algorithms for problems conjectured not to admit them, which supports his belief that P=NP. This is an opinion piece, not a resolution. https://cacm.acm.org/blogcacm/what-happened-in-fall-2026-ai-and-p-vs-np/ (search snippet only; the page was bot-blocked when I tried to fetch it) - ProofAtlas checked the status on 2026-08-07. It found a June 2026 arXiv preprint (Arthanari) claiming a Lean-verified P=NP proof and a July Lean package claiming P≠NP. Both declare axioms. Neither was audited, published in a qualifying outlet, or shown to have general acceptance. https://www.proofatlas.ai/collaboration/p-vs-np/ - Around 2026-09-22: Scott Aaronson says he has heard rumours that OpenAI and Anthropic may be withholding solved open problems. The rumours are secondhand and, per the report, are not about P vs NP. https://www.mindstudio.ai/blog/openai-anthropic-hiding-math-breakthroughs - About 2026-10-12: Aaronson said that if P=NP, it would be very amenable to AI because AI only needs to find an algorithm. He still strongly guesses P≠NP. https://www.developing.dev/p/e070-scott-aaronson - 2026-09: OpenAI presented a proposed counterexample for Navier–Stokes. Clay still lists that problem as "Active", with no prize awarded. This shows how labs and Clay handle claims. https://en.wikipedia.org/wiki/Millennium_Prize_Problems and https://mathworld.wolfram.com/MillenniumPrizeProblems.html BOTTOM LINE: The evidence favours No for now. No credible full resolution exists, and the fall-2026 AI results are explicitly short of P vs NP. The residual risk is a lab-produced, Lean-checked polynomial SAT algorithm or proof appearing before 2030, which is plausible given the pace of AI results but currently unevidenced. The contract's loose third route lowers the bar for a claim to count. NOT FOUND: I could not read Viola's full CACM post because it was bot-blocked. I found no result from the ETH authors' review of OpenAI's proof. I found no lab statement about attempting P vs NP. I did not look for the contract's own price or sibling markets.
claude-opus-5-5, high effort
38% conf
Yes 32%
No 68%
already resolved?
The condition is not met. Clay's P vs NP page, fetched 2026-11-20, still says "Unsolved". No lab or credible researcher has claimed a full resolution; the fall-2026 ETH refutations by OpenAI and Anthropic are explicitly short of P vs NP, and Brockman himself said P probably does not equal NP.
time left
About 3 years and 1 month remain before the Jan 1, 2030 deadline.
if nothing changes
If nothing changes, the market resolves No.
path to no
AI-lab results keep landing on adjacent questions (subexponential SAT, ETH and SETH variants), but polynomial-time SAT is never found and P≠NP stays blocked by the known relativization, natural-proofs and algebrization barriers. Any premature claim gets a publicly identified error, which triggers the contract's carve-outs.
path to yes
A lab publishes a Lean-verified polynomial SAT algorithm or a P≠NP proof, two Source Agencies report it as resolved, and no flaw is found before expiry. The contract's loose third route makes this possible well before any Clay recognition. The Kalshi anchor is a 53% mid, but the book is thin and wide (38/69 bid-ask, about 445 contracts a day), and the last daily close was 23%. That makes the mid a noisy signal rather than a strong consensus. The environment does justify far more than the historical base rate (Metaculus medians of 2054–2077): two labs independently refuted ETH with Lean proofs, there is a Navier–Stokes counterexample claim, the sibling Riemann market trades at 86%, and Aaronson reports labs racing on P vs NP. Even so, P vs NP is widely seen as the hardest of these problems, so I settle near 0.32, between the market's last close and its mid.
claude-fable-5-2, high effort
38% conf
Yes 40%
No 60%
already resolved?
The condition is not met: Clay's page (fetched 2026-11-20) lists P vs NP as "Unsolved", and Viola's Nov 17 post, Britannica and the follow-up research all confirm no lab or credible researcher claims a full resolution. The fall-2026 results (the ETH refutation by OpenAI and Anthropic, the Subercaseaux NTIME corollary) are explicitly short of P vs NP, and Brockman himself said "p still probably doesn't equal np".
time left
About 3.1 years remain until Jan 1, 2030.
if nothing changes
If nothing changes, this resolves No: nothing is scheduled, and the June/July 2026 Lean claims rest on axioms and have not been accepted.
path to no
P≠NP is true but the known proof barriers hold even against much stronger AI, or a claimed proof is reported and then has an error identified, which triggers the retraction/identified-error carve-outs as with Deolalikar in 2010.
path to yes
A frontier lab publishes a Lean-verified P≠NP proof or polynomial SAT algorithm in 2027-2029, and either two Source Agencies report it as resolved or it gets rapid peer-reviewed publication, both far faster than Clay's two-year wait. The pace supports this: a Navier-Stokes counterexample in September, ETH refuted by two labs in October with fast expert uptake, a Riemann claim today with Kalshi's Riemann-before-2027 market at 86%, and Aaronson describing the labs as racing on P vs NP. The Kalshi mid is 53% but thin and wide (bid 38 / ask 69, last daily close 23%, about 445 contracts a day), and it appears to have spiked on the Riemann news. Older Metaculus and Manifold forecasts of decades predate this run and are stale, but P vs NP is plausibly the hardest of the set and the believed direction is a lower-bound proof with no known strategy. I shade to about 0.40, near the bid.
4. Final plain mean of 2 draws, second round
36% (38% confidence)
Yes: 36% No: 64%
average of 2 draws (32%, 40%), redrawn after an extra research pass
already resolved?
The condition is not met. Clay's P vs NP page, fetched 2026-11-20, still says "Unsolved". No lab or credible researcher has claimed a full resolution; the fall-2026 ETH refutations by OpenAI and Anthropic are explicitly short of P vs NP, and Brockman himself said P probably does not equal NP.
time left
About 3 years and 1 month remain before the Jan 1, 2030 deadline.
if nothing changes
If nothing changes, the market resolves No.
path to no
AI-lab results keep landing on adjacent questions (subexponential SAT, ETH and SETH variants), but polynomial-time SAT is never found and P≠NP stays blocked by the known relativization, natural-proofs and algebrization barriers. Any premature claim gets a publicly identified error, which triggers the contract's carve-outs.
path to yes
A lab publishes a Lean-verified polynomial SAT algorithm or a P≠NP proof, two Source Agencies report it as resolved, and no flaw is found before expiry. The contract's loose third route makes this possible well before any Clay recognition. The Kalshi anchor is a 53% mid, but the book is thin and wide (38/69 bid-ask, about 445 contracts a day), and the last daily close was 23%. That makes the mid a noisy signal rather than a strong consensus. The environment does justify far more than the historical base rate (Metaculus medians of 2054–2077): two labs independently refuted ETH with Lean proofs, there is a Navier–Stokes counterexample claim, the sibling Riemann market trades at 86%, and Aaronson reports labs racing on P vs NP. Even so, P vs NP is widely seen as the hardest of these problems, so I settle near 0.32, between the market's last close and its mid.
Pipeline Timing and Cost
Total pipeline time: 110.6s
LLM cost: $1.66 (Sonnet 5.5 $0.91, Fable 5.2 $0.31, Opus 5.5 $0.22, 22 web searches $0.22)
Per-tool research timings shown in the Research section above.