When will AI solve the next Frontier Math problem?

• My median prediction for when AI will solve the next qualifying FrontierMath problem is 19th August 2026.
• Recent breakthroughs in May 2026, including OpenAI disproving the planar unit distance conjecture and Google DeepMind solving nine Erdős problems, demonstrate that AI models have already entered a period of high-intensity mathematical discovery.
• Documented progress on existing FrontierMath problems, such as GPT-5.4 Pro solving three of nine equations for the Diophantine finiteness problem, suggests that the “remaining” work for several problems is partially complete.
• The imminent rollout of next-generation reasoning models like GPT-5.5 and specialized “agentic” systems like Claude Mythos (released late May 2026) provides the necessary capability spike to tackle “Solid result” or “Major advance” tiers.
• Epoch AI’s scheduled workshops in June 2026 across global cities are likely to generate a surge in both new problems and verification attempts, leading to an official announcement by late summer.

Key figures

Historical context

The FrontierMath benchmark was launched in November 2024 by Epoch AI as a ‘stress test’ for expert-level mathematical reasoning. Historically, AI performance on mathematics has moved through phases: saturation of grade-school math (GSM8K) in 2023-2024, followed by rapid gains in competition-level math (AIME/IMO) in 2025. A key milestone occurred on March 23, 2026, when GPT-5.4 Pro achieved the first-ever solve of a FrontierMath ‘Open Problem’ (the Ramsey-style hypergraph problem). This followed years of specialized efforts, such as Google DeepMind’s FunSearch (2023) and AlphaProof (2024), which reached silver-medal IMO performance. The current era (2026) is defined by ‘agentic’ workflows where models like Claude Mythos use parallel sub-agents to verify their own reasoning, a technique that has historically preceded jumps in success rates on unsolved conjectures.

Tailwinds

• Model Iteration Cycle: The release of GPT-5.5 (April 2025) and Claude Opus 4.8 (May 2026) provides significantly more “thinking time” and better handling of “long-horizon” tasks necessary for research-level proofs.
• Agentic Architectures: New frameworks like Anthropic’s Claude Code and DeepMind’s Antigravity allow models to self-verify and iterate through proofs autonomously, reducing the human-in-the-loop bottleneck.
• Formal Verification Integration: The growth of the Lean mathlib library to over 210,000 theorems allows AI to check its own work against “ground truth” logic, drastically reducing hallucinations in complex proofs.
• Financial Momentum: R&D spending at Alphabet, Meta, and Microsoft reached historical peaks in early 2026, ensuring the high-compute environments required for FrontierMath are fully funded.

Headwinds

• Strict Verification Gaps: Epoch AI requires confirmed replication and often joint publication with the original problem contributor, which can delay official announcements by months even after a model produces a correct proof.
• Problem Depletion: If several “Moderately Interesting” problems are removed due to poor quality or existing literature (as happened on March 5, 2026), the remaining “Breakthrough” problems may represent a significantly higher difficulty wall.
• Restricted Access: Currently, only OpenAI is reported to have purchased access to the automated verifiers, potentially creating a bottleneck if other labs like Anthropic or DeepMind solve problems but lack the official pipeline for confirmation.
• Model Fragility: Despite high benchmark scores, models still show a 60% performance drop when “distractor” information is added, suggesting that some “solves” may be fragile or non-robust under Epoch’s stress-test conditions.

Detailed reasoning

My analysis of the FrontierMath benchmark indicates a high-velocity environment where the technical capability for a second solve likely already exists, but the official announcement is gated by the verification and publication protocols of Epoch AI. As of May 31, 2026, the benchmark shows only one official solve (a Ramsey-style hypergraph problem achieved by GPT-5.4 Pro in March). However, there is evidence of substantial “work-in-progress” that suggests a new announcement is imminent.

First, I have weighted the recent breakthroughs in Erdős conjectures heavily. In the final weeks of May 2026, OpenAI, Anthropic, and Google DeepMind all announced autonomous solutions to long-standing mathematical problems (notably the Planar Unit Distance Problem and nine other Erdős problems). While these may not be within the FrontierMath official set, the models achieving them—GPT-5.4 Pro, Claude Mythos, and Gemini 3.1 Pro—are the same systems currently targeting the benchmark. Historically, once a reasoning breakthrough occurs in a related field (like combinatorial geometry), official benchmark updates follow within 60 to 90 days as researchers refine proofs and close verification loops.

Second, the “partial progress” logged on the FrontierMath page is a strong leading indicator. GPT-5.4 Pro solved three of nine equations for the “Finiteness Problem for Diophantine Equations” as of March 5, 2026. Given the release of GPT-5.5 in late April 2026 and the subsequent increase in token generation speeds and reasoning efficiency, it is highly probable that the remaining components of such problems are being finalized now.

Third, the operational timeline of Epoch AI suggests a summer cluster of activity. Epoch AI scheduled workshops in New York, London, and Berkeley throughout June 2026 to expand the challenge set. Such workshops often serve as the deadline for labs to submit their latest results for verification to coincide with public announcements.

My median prediction of 19th August 2026 accounts for a “verification lag.” Even if a model solves a problem in June, the resolution criteria specify that we must wait for the official public update or announcement. Given that the first solve (March 23) took several weeks of verification by contributors like Will Brian, a 2-to-3-month window from the current flurry of May activity is the most statistically sound expectation. The high-confidence markets on Polymarket (82% for an IMO gold medal and Gemini scoring ≥40% by June 30) further support a summer “breakout” in AI reasoning.

Key uncertainties

• Epoch AI Publication Lag: The question resolves on the date of public announcement, not the date of the solve; administrative or editorial delays at Epoch AI could push the date into late 2026.
• Tier Difficulty Wall: While “Moderately Interesting” problems have been solved, the “Breakthrough” tier (0 of 3 solved) may require a paradigm shift in AI architecture that has not yet occurred.
• Verifier Access Bottleneck: If only one company maintains access to the verifiers, the lack of competition in the submission pipeline could slow the rate of official resolutions.

Conclusion

• The next FrontierMath solve is likely to occur between late June and September 2026, driven by the rollout of the GPT-5.5 series and competitive pressure from Anthropic and DeepMind.
• Decision-makers should monitor the Epoch AI dashboard following the conclusion of their June 2026 global workshops, as these events often serve as catalysts for benchmark updates and new problem submissions.
• The shift from general-purpose assistants to “agentic” reasoning systems (like Claude Mythos and Antigravity) is the primary technical driver behind the current acceleration in research-level problem solving.

Question: whether Epoch will publicly confirm a second FrontierMath: Open Problems AI solution by 2026-08-12 12:00 UTC. All teams agree the question resolves only on an Epoch public announcement (not internal model runs), and the strongest direct signal is negative: Epoch has announced one qualifying solve (a moderately interesting Ramsey-style hypergraph problem on 2026-03-23), then ~69 days without a second announcement as of 2026-05-31, while about 73 days remain. Forecasters emphasize positive but indirect capability signals—OpenAI has verifier access, FrontierMath benchmark scores and several external AI-math advances (OpenAI/Lean results, DeepMind reporting) show sharp improvement—and note partial progress on the Finiteness Diophantine task; however Epoch has already disqualified/removed at least one near-miss, showing that internal solves may not become qualifying announcements. Given the bottleneck of public verification and the remaining problems likely being harder than the first, the teams converge on a meaningful but minority chance (roughly one-third, ~30–33%) of a second public Epoch-confirmed solve by the deadline, with most remaining mass on no qualifying announcement. Updates to watch: Epoch page/Substack changes, OpenAI verifier/scaffold releases or statements, and concrete completion of the partially solved Diophantine subcases.

Forecast and Reasoning

Forecast: Probability distribution:

• 1.00% chance of value below 2026-06-03 12:00:00 UTC
• 5.00% chance of value below 2026-06-12 12:00:00 UTC
• 10.00% chance of value below 2026-06-23 12:00:00 UTC
• 20.00% chance of value below 2026-07-03 12:00:00 UTC
• 30.00% chance of value below 2026-07-17 12:00:00 UTC
• 40.00% chance of value below 2026-08-06 12:00:00 UTC
• 50.00% chance of value below 2026-08-14 12:00:00 UTC
• 60.00% chance of value below 2026-08-18 12:00:00 UTC
• 70.00% chance of value below 2026-08-21 12:00:00 UTC
• 80.00% chance of value below 2026-08-24 12:00:00 UTC
• 90.00% chance of value below 2026-08-27 12:00:00 UTC
• 95.00% chance of value below 2026-08-29 12:00:00 UTC
• 99.00% chance of value below 2026-08-30 12:00:00 UTC

Forecast Reasoning

We are forecasting the date of Epoch AI’s first public announcement/update confirming an AI solve of a new qualifying FrontierMath Open Problem (excluding the already-solved Ramsey-style hypergraph problem). As of 2026-05-31, only one solve has been publicly confirmed on the Open Problems page, and there has been a ~2+ month gap since the March confirmation without another announced solve. That pushes toward a relatively low short-horizon base rate, especially since remaining problems skew harder.

Countervailing evidence is that (per multiple analysts’ review of the problem pages/posts) there are signs of active progress on at least one other “moderately interesting” problem (notably the book-graphs/off-diagonal Ramsey-related problem) and model capability has continued to advance. However, the resolution depends on Epoch’s public confirmation timing, which can lag discovery (verification + contributor confirmation + site update), and verifier access appears constrained, adding additional delay risk.

Synthesizing these, I place the probability of a qualifying Epoch announcement/update occurring on or before 2026-08-12T12:00Z at roughly the mid-to-high 40%s, with the median falling just after the deadline (so >50% in the open-after tail). Within-window mass is concentrated in July and early August rather than early June, reflecting both the observed announcement cadence and verification latency uncertainty.

An amateur forecast would commonly (a) over-extrapolate from the single March solve to assume another quick solve is imminent, or (b) anchor on the question deadline and incorrectly pile probability at 2026-08-12 rather than using the open-after tail. This forecast instead blends the observed low realized announcement rate since launch with credible but uncertain near-term progress signals, and explicitly represents substantial “no in-window announcement” risk by placing the median and higher percentiles strictly past the deadline.

Forecast rationale (numeric):

— Iteration 1 — Across the forecasts, the core reasoning is built around a few shared drivers:

Main factors driving the predictions

• A first solve has already happened, which is strong evidence that the benchmark is now within reach of frontier models.
• Rapid model progress matters, especially if upcoming releases or hidden internal capabilities can extend FrontierMath performance.
• Verification/public announcement is a bottleneck: a solve may occur before it is publicly confirmed, and Epoch’s review process can add delay.
• The remaining problems may be harder: several rationales note a selection effect, where easier problems were likely solved first, leaving a tougher tail.
• Access is limited: verifier access appears concentrated in a small number of labs, which constrains who can demonstrate and validate a solve.

Areas of consensus

• All forecasts treat a near-term solve as plausible, not implausible.
• All assume the relevant event is public confirmation, not just a private internal solve.
• All incorporate a deadline-aware tail, acknowledging that if it does not happen soon, the next solve could slip much later.

Main disagreements

• Timing confidence varies sharply:
  • One forecast is quite optimistic, centering on late June to early July 2026.
  • Another is moderately optimistic, with a center around late July 2026 and substantial probability before Aug. 12, 2026.
  • The most cautious forecast puts most probability after the deadline, with a long tail into late 2026–2027.
• The optimistic side leans more on benchmark momentum and expected near-term breakthroughs, while the cautious side emphasizes harder remaining tasks and process delays.

Overall synthesis

The forecasts agree that the next FrontierMath solve is driven by a mix of frontier model progress, benchmark momentum, and verification constraints. The key uncertainty is not whether AI can eventually solve more problems, but how quickly the next publicly confirmed solve will appear, given that the remaining tasks may be substantially harder and that announcement timing depends on external validation.

— Iteration 2 — Across the forecasts, the main story is a combination of fast-improving AI math capability and a lot of event-specific friction around verification and publication.

Shared reasoning patterns

• Capability is improving quickly. All the rationales point to recent progress in AI math, including benchmark gains and at least one FrontierMath Open Problem already being solved.
• The bottleneck is not just solving, but proving it publicly. A recurring theme is that the relevant event depends on Epoch’s verification and announcement process, not merely whether a model privately finds a solution.
• External factors matter. Access to verifiers, publication standards, and lab disclosure incentives all shape when a solve becomes visible.
• Uncertainty remains high. Even optimistic takes keep wide tails because the remaining problems may be much harder than expected, and the timing of publication could lag behind the actual solve.

Areas of consensus

• There is broad agreement that AI is close enough to make a solve plausible in the near term, rather than being years away.
• There is also agreement that late summer 2026 is a plausible window if the event happens soon.
• The rationales all treat public confirmation as the key milestone, so timing reflects both technical progress and administrative delay.

Main disagreement

• The biggest split is whether the announcement happens before the Aug. 12, 2026 deadline:
  • One view is highly optimistic, putting substantial probability on a July 2026 announcement.
  • Another is cautiously optimistic, but still says the most likely outcome is after the deadline.
  • A third is more moderate, assigning only about one-third chance before the deadline and centering the median in fall 2026.
• So the divergence is less about whether a solve will happen eventually, and more about how soon public confirmation will arrive.

Bottom-line synthesis

The collective reasoning suggests that AI math ability is advancing fast enough that a FrontierMath solve is increasingly plausible soon, but the event’s timing is constrained by verification and publication delays. The central uncertainty is not raw capability; it is whether a qualifying, publicly confirmed announcement will arrive before the August deadline or slip into late 2026.

— Iteration 3 — Across the forecasts, the main view is that the next FrontierMath solve is plausibly near-term, with a central estimate in mid-to-late July 2026, and with a meaningful chance it will be publicly announced before the Aug. 12, 2026 cutoff. One forecast is relatively bullish on timing, another is more cautious, but they agree that the most likely window is summer 2026, not far in the future.

Main reasoning patterns

• Recent progress is a strong signal.
  The first FrontierMath solve in March 2026, plus other recent high-profile math achievements, is taken as evidence that frontier math capability is improving quickly.
• OpenAI’s verifier access matters.
  Since OpenAI has exclusive access to the FrontierMath verifiers, the forecasts treat it as the most likely actor to produce another qualifying solve.
• Announcement timing is distinct from solve timing.
  Even if a solution already exists or is close, Epoch’s verification and announcement process can create a lag, so the “solve date” and “public announcement date” may not match exactly.
• The base rate is still limited.
  Only one problem has been solved so far, so the evidence base is thin and confidence remains moderate rather than high.
• Harder remaining problems may slow progress.
  There is concern that the easiest problems may already be gone, leaving tougher ones that could take much longer.
• Long tails are important.
  All forecasts allow for substantial uncertainty: a solution could come very soon if work is already in progress, but there is also a significant possibility of delays stretching into late 2026 or 2027.

Areas of agreement

• Another FrontierMath solve is more likely than not to happen eventually.
• The most plausible near-term window is July 2026, with some chance of an earlier June breakthrough.
• Verification and publication delays are a major source of uncertainty.

Areas of disagreement

• The forecasts differ mainly on confidence level: one puts a relatively high probability on a pre-deadline announcement, while another is notably more conservative.
• They also differ on how much weight to give the limited historical base rate versus the rapid pace of recent AI math progress.

Bottom line

The collective reasoning points to a near-term but uncertain breakthrough, centered around mid-to-late July 2026, driven by rapid AI math advances and OpenAI’s verifier access, but tempered by the small number of prior successes and the possibility that remaining FrontierMath problems are substantially harder.

Synthesis of Rationales

Accelerating Base Rates and Model Capabilities Forecasters broadly agree that the interval between AI mathematical breakthroughs is rapidly compressing. Following the successful solution of the first FrontierMath open problem in early 2026, a surge of advanced models released in May 2026 (including OpenAI’s GPT-5.5 ecosystem, DeepMind’s AlphaProof Nexus, and new Claude iterations) has introduced powerful agentic reasoning and scaled test-time compute. These neuro-symbolic systems are highly capable of tackling the benchmark.

Vulnerability of Remaining Problems The remaining FrontierMath problems include combinatorial and constructive challenges (such as Ramsey-style graphs and Hadamard matrices) that are highly susceptible to the strengths of current models. Because these problems can be attacked via massive programmatic brute-force searches and Python scripting, they do not necessarily require profound conceptual leaps. Furthermore, upcoming Epoch AI workshops intended to source new problems may introduce additional, highly tractable “low-hanging fruit.”

Verification Pipelines and Catalysts While Epoch AI can programmatically verify constructive solutions very quickly, formal announcements require rigorous human peer review, which introduces a natural delay of a few weeks between a breakthrough and its publication. Forecasters note that the summer months provide strong catalysts for an announcement, driven by targeted model testing, dataset error-correction updates, and major academic conferences (like ICML and IMO) that serve as prime targets for AI lab PR.

Bottlenecks and Stagnation Risks The primary argument against an imminent breakthrough rests on a potential “capability wall.” If the remaining unsolved problems resist current test-time compute and combinatorial search techniques—requiring deep spatial reasoning, abstract leaps, or facing massive formalization hurdles—the current generation of AI could stall. In this scenario, a solution would be delayed until fundamentally new model architectures are released in late 2026 or beyond.

SUMMARY

Question: When will AI solve the next Frontier Math problem? Final Prediction: Probability distribution:

• 10.00% chance of value below 2026-07-01 00:00:00 UTC
• 20.00% chance of value below 2026-07-16 08:00:00 UTC
• 40.00% chance of value below 2026-08-08 16:00:00 UTC
• 60.00% chance of value below 2026-10-22 00:00:00 UTC
• 80.00% chance of value below 2027-06-03 08:00:00 UTC
• 90.00% chance of value below 2028-06-09 16:00:00 UTC

Total Cost: extra_metadata_in_explanation is disabled Time Spent: extra_metadata_in_explanation is disabled LLMs: extra_metadata_in_explanation is disabled Bot Name: extra_metadata_in_explanation is disabled

Report 1 Summary

Forecasts

Forecaster 1: Probability distribution:

• 10.00% chance of value below 2026-06-30 00:00:00 UTC
• 20.00% chance of value below 2026-07-17 00:00:00 UTC
• 40.00% chance of value below 2026-08-15 00:00:00 UTC
• 60.00% chance of value below 2026-11-05 00:00:00 UTC
• 80.00% chance of value below 2027-03-25 00:00:00 UTC
• 90.00% chance of value below 2027-10-15 00:00:00 UTC

Forecaster 2: Probability distribution:

• 10.00% chance of value below 2026-07-02 00:00:00 UTC
• 20.00% chance of value below 2026-07-16 00:00:00 UTC
• 40.00% chance of value below 2026-08-06 00:00:00 UTC
• 60.00% chance of value below 2026-10-05 00:00:00 UTC
• 80.00% chance of value below 2027-04-15 00:00:00 UTC
• 90.00% chance of value below 2028-01-15 00:00:00 UTC

Forecaster 3: Probability distribution:

• 10.00% chance of value below 2026-07-01 00:00:00 UTC
• 20.00% chance of value below 2026-07-16 00:00:00 UTC
• 40.00% chance of value below 2026-08-05 00:00:00 UTC
• 60.00% chance of value below 2026-10-25 00:00:00 UTC
• 80.00% chance of value below 2027-10-01 00:00:00 UTC
• 90.00% chance of value below 2029-06-30 00:00:00 UTC

Research Summary

The research summarizes progress on Epoch AI’s FrontierMath benchmarks and recent AI advances in mathematical research up to May 31, 2026. FrontierMath: Open Problems lists 15 curated research-level problems; one (a Ramsey-style hypergraph problem) was solved by GPT-5.4 Pro in March 2026, verified by the problem contributor and formalized in Lean (~6,300 lines), and Epoch AI publicly announced the solved problem on May 31, 2026. FrontierMath’s broader Tier 4 benchmark shows substantially higher AI performance (e.g., DeepMind’s AI Co‑Mathematician reached 48% on Tier 4, GPT-5.5 Pro ~39.6%, GPT-5.4 Pro ~37.5%, Claude Opus ~22.9%), but the Open Problems subset is smaller (15 problems) and oriented toward harder, research-grade tasks; remaining Open Problems span number theory, combinatorics, algebraic geometry, and topology/geometry across difficulty categories from “moderately interesting” to “breakthrough.”

The research also documents relevant meta-facts: agentic/multi-component systems (e.g., AI Co‑Mathematician) outperform single models on research problems; recent model releases (GPT-5.5, Claude Opus updates, DeepMind systems) drove rapid capability gains; verification and announcement lags (the March solve was announced in late May) and human-in-the-loop work were integral to the confirmed solve; no active prediction markets specific to FrontierMath solves were found on Manifold, Metaculus, Polymarket, or similar platforms; and the empirical base rate to date is one Open Problem solved in the first ~5 months after the benchmark’s launch, while Tier-4 solve rates show faster progress on broader sets of problems than on the curated Open Problems.

Sources/websites searched or cited (domains provided):
https://epoch.ai
https://github.com
https://twitter.com (X)
https://deepmind.com
https://openai.com
https://manifold.markets
https://metaculus.com
https://polymarket.com
https://kalshi.com
(Also referenced: the Soohak research-level benchmark and various news coverage and model release notes from April–May 2026.)

RESEARCH

Report 1 Research

I’ll help you gather comprehensive information about FrontierMath, AI mathematical capabilities, prediction markets, and relevant base rates. Let me conduct several searches to build a detailed research base.

Forecast — Post 389

Modal: Aug 12, 2026 to Oct 24, 2026 (20.0%) • partial_shift • 6m18s

Interpretation

Reading: strict

The resolution criteria are highly specific and literal: they key resolution strictly to the Epoch AI public announcement or page-update timestamp (not the solve date), explicitly exclude the already-solved hypergraph problem, define a precise annulment condition for timestamp ambiguity, and specify an exact UTC deadline — all of which point to a narrow, literal reading with little room for charitable interpretation.

Time bounds

• observation_window: Any qualifying public announcement or update to the FrontierMath Open Problems page by Epoch AI on or before 2026-08-12T12:00:00+00:00
• resolution_date: 2026-08-12

Edge cases identified

1. The Ramsey-style hypergraph problem is explicitly excluded; only new solves of previously unsolved problems qualify.
2. Resolution is keyed to Epoch AI’s public announcement or page update, not the underlying solve date — a solve that occurs before the deadline but is announced after does NOT qualify.
3. If multiple problems are solved simultaneously and announced together, the single announcement datetime is used.
4. Annulment applies only to the ambiguous-precision scenario (announcement date straddles the deadline with insufficient precision); a clear announcement after the deadline simply resolves as ‘>2026-08-12T12:00:00+00:00’.
5. The fallback to ‘other reliable public reports’ of Epoch AI confirmation could create ambiguity about what counts as an authoritative source if the Epoch AI site is down.
6. It is unclear whether a partial announcement (e.g., ‘a problem has been solved’ without specifying which problem) qualifies if the problem’s identity is confirmed only later.
7. The phrase ‘qualifying FrontierMath problem’ relies on the official FrontierMath set as defined by Epoch AI; if the set is expanded or redefined before the deadline, it could affect what counts as qualifying.

Research (3/3 variants, shared evidence pool)

Total evidence registered (shared pool): ?

Research Brief

Evidence confidence: medium

Scenario 1: Qualifying announcement before August 12, 2026 deadline — mid-July timing [medium evidence]

Conditions favoring

Accelerating AI mathematical capabilities (per sources 13, 21, 39); near-progress on Hadamard Matrices (high author-estimated solvability, per source 36) and Ramsey Numbers for Book Graphs (partial AI progress to n=56, per source 35); rising-hazard survival model P≈72% with median mid-July (per source 26); Manifold market at ~80% for 2+ solves before August (per source 8); multiple frontier models now capable of solving the first problem (per source 29); Epoch AI actively running FM:OP workshops that may stimulate problem engagement (per source 19).

Conditions against

69+ days since first solve with no second announcement despite active frontier model testing (per source 9, 24); remaining problems explicitly designed to require publishable-result-level difficulty (per source 15); first problem was described as potentially ‘low-hanging fruit’ relative to others (per source 10); Epoch AI’s own scaffold testing of GPT-5.4 Pro on other problems found no solves (per source 9).

Scenario 2: No qualifying announcement by August 12, 2026 — question resolves ‘>2026-08-12’ [medium evidence]

Conditions favoring

69+ days of silence post-first-solve (per sources 22, 24, 30); remaining 14 problems rated Solid Result, Major Advance, or Breakthrough difficulty (per source 4, 16); Epoch AI’s own scaffold found GPT-5.4 Pro solved no other Open Problems (per sources 9, 20); conservative hazard model gives P≈48% for solve by Aug 12, meaning >50% mass on no-solve (per source 18); problems require ‘publishable result’ bar that previously led to one problem removal (per source 15, 31); Epoch AI in quality-review and workshop-development mode rather than active solve evaluation (per sources 2, 14, 19).

Conditions against

Rising AI capabilities trajectory; multiple frontier models now capable of the first solve (per source 29); Manifold market at 80% for 2+ solves before August (per source 8); near-progress noted on at least two ‘Moderately Interesting’ problems (per sources 35, 36); 10+ weeks remain until deadline from May 31.

Scenario 3: Qualifying announcement before August 12, 2026 — near-term (before July 2026), driven by Hadamard Matrices or Ramsey Book Graphs [low evidence]

Conditions favoring

Hadamard Matrices problem has high estimated solvability (95-99% per source 36) and is rated ‘Moderately Interesting’; Ramsey Numbers for Book Graphs has active partial AI progress (n≤56 constructions found, per source 35); Manifold market at 44% for announcement before July (per source 8); workshops may accelerate problem engagement (per source 19); GPT-5.5 and newer models may outperform the GPT-5.4 Pro that failed in March 2026.

Conditions against

Hadamard Matrices problem remains officially unsolved despite high estimated solvability; n=100 Ramsey challenge was just raised, suggesting full solve is further away than completion of n≤56 implies (per source 35); no announcement in the 10 weeks post-first-solve despite the most capable models being available (per sources 9, 22); Epoch AI workshops are focused on creating new problems, not solving existing ones (per source 19).

Scenario 4: Announcement occurs but is annulled due to timestamp ambiguity around the August 12 deadline [low evidence]

Conditions favoring

Annulment risk explicitly noted in resolution criteria; if Epoch AI announces a solve in early-mid August 2026 without a precise timestamp, the question could be annulled rather than resolving as ‘>2026-08-12’; the conservative survival model’s median of ~mid-August (per source 18) places a non-trivial mass of scenarios near the deadline where ambiguity could arise.

Conditions against

Epoch AI’s prior announcement (March 23, 2026) included a precise Substack post with datestamp (per source 3), suggesting Epoch AI communications are typically timestamped; the FrontierMath Open Problems page updates have been logged with precise dates (per sources 6, 31); annulment requires both an announcement and insufficient timestamp precision, which is a narrow conjunction.

Key uncertainties

• Whether the remaining 14 unsolved FrontierMath Open Problems are substantially harder than the first (Moderately Interesting hypergraph problem), which would extend solve timelines significantly.
• Whether any frontier model (GPT-5.5, future releases, or Gemini/Claude variants) has attempted all remaining Open Problems in a scaffold post-March 2026 and whether results exist but have not been publicly announced.
• The pace at which AI mathematical capabilities are improving — whether the step from GPT-5.4 to GPT-5.5/future models is sufficient to tackle Solid Result, Major Advance, or Breakthrough tier problems.
• Whether Epoch AI’s FM:OP workshops (late May - June 2026) or other institutional activities will accelerate or delay official evaluation and announcement of new solves.
• The ‘publishable result’ bar requirement: whether partial AI progress (e.g., Ramsey Book Graphs n≤56) will eventually be deemed sufficient for an official solve announcement.
• Whether the Hadamard Matrices problem (high expert solvability, Moderately Interesting) will be targeted and successfully solved given its estimated 1-4 week human expert solve time.

Forecast variants (3/3)

Variant 0: inside_view (inside_view_v1)

Inside View — Causal Mechanism Analysis

Current state → Decision → Outcome chain:

1. Current state: As of May 31, 2026, exactly 1 of 15 FrontierMath Open Problems has been solved (announced March 23, 2026). There have been 69+ days of silence post-first-solve with no second announcement (per sources 22, 24, 30). The first solved problem was “low-hanging fruit” — computational combinatorics, ~250k tokens, not traditional proof (per source 10).

2. Causal mechanisms pointing toward a solve within the window:

   • AI capabilities are accelerating rapidly: GPT-5.5 Pro solved PhD-thesis-level additive number theory open problems in under 2 hours (per source 39); DeepMind’s AI Co-Mathematician hit 48% on FrontierMath Tier 4 (per source 13)
   • Hadamard Matrices (Moderately Interesting) has 95-99% estimated solvability within 1-4 weeks for an expert human (per source 36) — this is a strong causal signal that a sufficiently capable AI should be able to solve it soon
   • Ramsey Numbers for Book Graphs saw AI find constructions for all n≤56 (per source 35) — substantial partial progress
   • Multiple frontier models now capable of solving the first problem: GPT-5.4 (xhigh), Gemini 3.1 Pro, Claude Opus 4.6 (max) (per source 29)

3. Causal mechanisms pointing against a solve within the window:

   • Epoch AI’s own scaffold testing of GPT-5.4 Pro found ZERO other solves as of March 2026 (per sources 9, 20)
   • The 69+ day silence is itself a strong signal: if Hadamard Matrices were easily solvable by current models, it would likely have been solved already given the Epoch AI scaffolding apparatus
   • The “publishable result” bar is high — it previously led to one problem removal (per sources 15, 31)
   • Epoch AI is in workshop-development mode (per sources 14, 19) rather than actively running evaluations, possibly slowing announcement cadence

Quantitative anchors:

• Rising-hazard survival model (shape=1.5): P(solve by Aug 12) ≈ 71.6%, median ~mid-July (per source 26) — relative_evidence_strength: medium
• Conservative survival model (shape=0.7): P ≈ 47.8%, median ~mid-August (per source 18) — relative_evidence_strength: medium
• Manifold Markets: ~80% for 2+ solves before August, 44% before July, 50% before August (per source 8) — aggregate crowd signal
• The two survival models bracket: ~48-72% probability of solve within window

My inside-view synthesis: The 69-day silence post-first-solve is the strongest single causal signal. If the Hadamard Matrices problem (author-estimated 95-99% solvable by expert in 1-4 weeks) hasn’t been solved despite active attention, it suggests current models are not quite there yet for even the “easiest” remaining problems. However, AI capabilities are clearly improving rapidly (GPT-5.5, newer models), and the Hadamard Matrices problem remains the most plausible near-term candidate. I weight the solve-within-window probability at roughly 55-60% — splitting between the two survival models, slightly discounting for the silence signal and the harder remaining problems.

Timing distribution within the window (conditional on solve occurring):

• If a solve happens, the most likely problems are Hadamard Matrices or Ramsey Book Graphs (each ~26% per Manifold, per source 40)
• The Epoch AI workshops (May-June) may stimulate engagement but are focused on new problems, not existing ones
• A mid-June to mid-August announcement window seems most likely if it happens
• I spread mass roughly uniformly across the bins within the window, with slightly more weight toward July-August (as the silence so far implies it’s not imminent)

Bin allocation strategy:

• P(no solve by Aug 12) ≈ 40% → bin_8+ (the “>deadline” bin and beyond)
• P(solve before Aug 12) ≈ 58% spread across bins 0-7
• I heavily discount May 31 - Jun 9 (bin_0) since 69 days of silence argues against immediate solve
• Bins 1-3 (Jun 9 - Jul 7): low-medium, workshops ending might prompt evaluations
• Bins 4-7 (Jul 7 - Aug 12): higher, as this is when a mid-July to early August solve would be announced
• The “after deadline” mass: the conservative model puts median at mid-August, so substantial mass should be in bins 8-12
• Annulment risk (bin scenario 4, low evidence strength) is negligible given Epoch AI’s track record of precise timestamps

Distribution:

• May 31 - Jun 9 (bin_0): 3% — very unlikely given 69-day silence
• Jun 9 - Jun 18 (bin_1): 6%
• Jun 18 - Jun 27 (bin_2): 7%
• Jun 27 - Jul 7 (bin_3): 9%
• Jul 7 - Jul 16 (bin_4): 10%
• Jul 16 - Jul 25 (bin_5): 10%
• Jul 25 - Aug 3 (bin_6): 8%
• Aug 3 - Aug 12 (bin_7): 5% — close to deadline, strict timestamp rules
• Aug 12 - Oct 24 (bin_8): 15% — solve happens but announced just after deadline, or in near future
• Oct 24 - Jan 5, 2027 (bin_9): 10%
• Jan 5 - Mar 18, 2027 (bin_10): 8%
• Mar 18 - May 30, 2027 (bin_11): 5%
• After May 30, 2027 (bin_12): 4%

Sum: 3+6+7+9+10+10+8+5+15+10+8+5+4 = 100% ✓

The large bin_8 reflects that the conservative model puts its median right at mid-August, meaning many scenarios involve an announcement shortly after the deadline. The mass on bins 9-12 reflects the possibility that the remaining problems are substantially harder and may take many more months.

Variant 1: outside_view (outside_view_v1)

Reference class: “Announcements of qualifying FrontierMath Open Problem solves by Epoch AI after the first announced solve.” Historical frequency is uninformative (n=1): the benchmark produced one announcement (Mar 23, 2026) ~2 months after the pilot release, but sample size is too small to form a reliable empirical base rate (see source 3, source 30). Adjustment: the brief contains two formal time-to-event / survival projections that bracket plausible outcomes — a rising-hazard model giving P(by Aug 12) ≈ 71.6% with median mid‑July (source 26) and a conservative constant‑hazard model giving P(by Aug 12) ≈ 47.8% with median ≈ mid‑August (source 18). I treat those models as strong quantitative evidence (per the brief’s instructions) and form an outside‑view aggregate near their midpoint while widening for the substantial epistemic uncertainty (69+ days of silence since the first solve, source 22; plus the remaining problems span harder tiers, source 4). Final: I place ~60% total probability that Epoch AI publicly announces a qualifying FrontierMath solve on or before 2026-08-12T12:00:00Z (bins bin_0–bin_7 sum to 0.60), and ~40% that no such announcement occurs by the deadline (bins bin_8–bin_12 sum to 0.40). Distributional shape: mass is concentrated in mid‑July (bins bin_4–bin_5) reflecting the rising‑hazard model’s mid‑July median and marketplace signals (Manifold markets showing substantial odds for another solve before August, source 8), with nonzero tail into early June and early August to reflect both faster breakthroughs (Hadamard/Book‑graph partial progress, sources 36 and 35) and the possibility announcements slip toward the deadline. I allocated most post‑deadline mass to the earliest after‑deadline bin (bin_8 = Aug 12–Oct 24, 0.30) because the conservative model’s median is near mid‑August (source 18) and workshop / review activity (sources 14, 19, 2) could delay public confirmation until shortly after the deadline. Key evidentiary anchors cited inline: no second announcement as of May 31 (sources 22, 30), rising‑hazard and conservative survival models (sources 26 and 18), marketplace signals (source 8), partial problem‑specific progress (Hadamard and Ramsey Book Graphs, sources 36 and 35), and Epoch AI’s stated publishable‑result bar and prior removal (source 15). Reading: strict — resolution depends on a verifiable Epoch AI public announcement or page update timestamp; annulment risk exists but is judged low relative to the main outcomes (the resolution rules require precise timestamping; if Epoch’s announcement lacks sufficient timestamp precision the question would be annulled). This forecast is the outside‑view aggregate (I avoided detailed inside‑view adjustments beyond the midpoint of the survival models and widened tails to reflect the limited historical sample and active uncertainties cited above).

Variant 2: contrarian_v2 (contrarian_v2)

Adversarial reading

The implicit consensus in the brief is that accelerating frontier-model capabilities plus partial progress on two Moderately Interesting problems will produce a qualifying Epoch AI announcement sometime between mid-July and mid-August 2026, with the bulk of probability mass before the 2026-08-12 deadline.

Counter-evidence found (or absence)

The 69-day silence after the March 23 announcement (sources 22, 24, 30) is substantive counter-evidence: GPT-5.4 Pro itself solved no other Open Problems in Epoch’s scaffold (sources 9, 20), the first solved problem was described as low-hanging computational combinatorics (source 10), and the remaining 14 problems sit in Solid Result / Major Advance / Breakthrough tiers (source 4). Epoch is currently running quality reviews and new-problem workshops rather than active solve evaluations (sources 2, 14, 19). This pattern contradicts the optimistic rising-hazard model and the Manifold 80 % “two solves by August” market.

Distribution implications

Because the 69-day gap and tier-difficulty evidence outweigh the accelerating-capability anecdotes, I shift mass away from the early-July bins (0–4) and toward the post-deadline terminal bin (bin_12) and the Aug 12–Oct window (bin_8). The near-term “Hadamard or Ramsey Book Graphs by July” scenario receives only low-evidence weight, so bins 0–3 are kept minimal. The resulting distribution is materially later than the brief’s consensus midpoint.

Consolidate

Strategy: logit_mean • Drafts: 3/3

Reviewers (3/3)

anchoring → shift_higher (low magnitude, medium confidence) anchoring_v1

The draft forecast’s central mass (p50) sits in the Jul 07–Jul 25 range, anchored primarily by the rising-hazard survival model’s “median mid-July” estimate (source 26). However, this rising-hazard model (shape=1.5) assumes an INCREASING solve rate over time — a heroic assumption given that 69+ days have already elapsed since the first solve with zero second announcement, and Epoch AI’s own scaffold testing found GPT-5.4 Pro solved no other Open Problems (sources 9, 20). The brief explicitly tags both survival models as “medium” evidence strength, meaning neither should dominate; a flat-prior reading of the two models brackets P(solve by Aug 12) at 48–72%, implying the modal scenario should sit closer to mid-August or post-deadline, not mid-July. The draft places ~58% mass before the deadline and uses the optimistic model’s median as its anchor, despite the 69-day silence being a direct disconfirmation of the rising-hazard assumption. A median-expectation forecast should weight the two models more symmetrically, shifting the p50 to approximately late July or early August rather than mid-July, and placing somewhat more mass in post-deadline, near-term (bin_8).

Flagged concerns

• Optimistic model anchoring: The draft’s modal bin is Jul 07–Jul 25 (bins 4–5, combined ~20.5%), reflecting the rising-hazard model’s mid-July median. But this model assumes shape=1.5 (accelerating solve rate), which is directly contradicted by 69+ days of silence post-first-solve. A flat-prior synthesis of both survival models would yield a p50 closer to late July or early August.
• Premature early-window mass: Bins 0–3 (May 31 – Jul 7) carry ~17% probability. Given 69 days of silence already elapsed without a solve despite active frontier model availability (source 9), placing 17% mass in the first ~5 weeks seems elevated relative to what either survival model implies for the near-term hazard.
• Manifold market over-reliance: The rationale cites the Manifold ‘~80% for 2+ solves before August’ market as supporting the solve-within-window scenario, but Manifold markets in this domain can be driven by AI-enthusiast sentiment rather than calibrated base rates. The 69-day empirical silence is a harder fact than crowdsourced market sentiment.

ceiling → shift_lower (medium magnitude, medium confidence) ceiling_v1

The key structural constraint the draft underweights is the “publishable result” bar imposed by Epoch AI, combined with the explicit evidence that the 69-day silence post-first-solve occurred despite active frontier model testing. Source 10 explicitly labels the first solved problem as “low-hanging fruit” — computational combinatorics rather than traditional proof — while the remaining 14 problems span Solid Result (5), Major Advance (3), and Breakthrough (3) tiers alongside 3 remaining Moderately Interesting problems. Critically, Epoch AI’s own scaffold testing of GPT-5.4 Pro (source 9, 20) found zero other solves across all remaining Open Problems as of March 2026, establishing a hard empirical floor: even the most capable model at the time of the first solve could not immediately cascade to a second solve. The draft places ~58% probability inside the window (bins 0–7), but this fails to integrate the structural constraint that the Ramsey Book Graphs problem was explicitly NOT credited as solved despite AI finding constructions for n≤56 (source 35) — Epoch AI raised the challenge to n=100, demonstrating the publishable-result ceiling actively rejecting partial AI progress. This constraint should push significantly more mass toward the post-deadline bins (bin_8 through bin_12).

Flagged concerns

• Publishable-result ceiling suppressing qualifying announcements: Epoch AI’s ‘publishable result’ bar (source 15, 31) functioned as a structural ceiling that already prevented the Ramsey Book Graphs partial solve (n≤56 constructions, source 35) from qualifying, and previously removed ‘Explicit Deformations of Algebras’ entirely. This ceiling is active and constrains the effective solve rate below the raw capability rate — the draft treats AI capability improvements as directly translating to qualifying announcements, but the institutional filter adds a substantial delay/rejection layer the distribution doesn’t reflect.
• Hard empirical zero: GPT-5.4 Pro solved no other problems in scaffold testing: Source 9 and 20 establish that Epoch AI’s scaffold testing of GPT-5.4 Pro — the same model that solved the first problem — found zero qualifying solutions across all remaining Open Problems. This is a hard constraint: as of March 2026, no other problem was solvable even by the best available model. The 69-day silence through May 31 confirms this. The draft places 3% in bin_0 and 6% in combined ~9% for the first 18 days (bin_1) but given that newer models like GPT-5.5 would need to be specifically evaluated by Epoch AI in scaffold testing — a process with institutional latency — the near-term bins should be near-zero.
• Institutional bottleneck: Epoch AI in workshop/review mode, not evaluation mode: Sources 2, 14, and 19 confirm Epoch AI is conducting FM:OP workshops to develop NEW problems (not solve existing ones) and running a Tiers 1-4 quality review. This institutional posture creates a structural delay: Epoch AI must actively run scaffold evaluations and verify results before announcing. The draft does not account for this organizational ceiling on announcement cadence — the rising-hazard model (source 26) assumes a mechanical solve rate without modeling Epoch AI’s capacity/willingness to run evaluations.

math → redistribute (medium magnitude, medium confidence) math_v1

The draft forecast sums to 1.00 (verified to machine precision) and contains no arithmetic errors in the published distribution. However, there is a critical inconsistency between the stated inside-view reasoning and the final bin allocation: the inside-view explicitly states “I weight the solve-within-window probability at roughly 55-60%” and allocates bins 0-7 to 58%, yet the contrarian variant explicitly shifts that downward “away from the early-July bins (0–4) and toward the post-deadline terminal bin (bin_12).” The logit_mean aggregation of three variants should have transparently resolved this conflict, but the output shows no derivation of how 55-60% was merged with the contrarian downward shift. This leaves the forecast’s coherence unexplained: is P(solve by Aug 12) actually 58%, or is the contrarian view’s heavier post-deadline mass pulling it lower? A checksum of published bin_0–accounting for label mapping (bin_7) yields approximately 0.485, which suggests the final distribution has drifted below even the consensus 55-60% and toward the contrarian 40-45%, yet no reconciliation of these three branches is visible in the rationale.

Flagged concerns

• Ensemble reconciliation unexplained: The rationale presents three separate forecasting variants (inside_view ~55-60%, outside_view ~60%, contrarian_v2 ~40-45%), states they are ‘consolidated via logit_mean,’ but provides no intermediate calculation showing how logit_mean aggregation of those percentages produced the final P(before Aug 12). The published distribution (bins 0–7) sums to approximately 48–50%, which appears to be below the stated inside-view 58% and closer to the contrarian 40–45%, yet there is no transparent explanation for the weighting of the three branches.
• Conditional timing distribution not documented: The inside-view rationale states ‘I spread mass roughly uniformly across the bins within the window, with slightly more weight toward July–August,’ but the actual published distribution (bins 0–7) shows a modal region around bins 4–5 (Jul 7–Jul 25) with lighter tails in June and early August. The phrase ‘roughly uniformly’ does not match a distribution with ~10% in each of bins 4–5 and ~6–9% in earlier bins. The distribution is defensible, but the stated shape and the actual shape are not explicitly reconciled.
• Post-deadline terminal bin allocation inconsistent with survival model medians: The inside-view rationale cites the conservative survival model’s median of ~mid-August (source 18) as justification for allocating substantial mass to bins 8–12. However, if the median remaining wait is ~78 days from May 31, that places the median announcement ~mid-August, which is AFTER the deadline (Aug 12). For a distribution whose median is mid-August, approximately 50% of the total probability mass should be in bins 0–7 and ~50% in bins 8–12. The published distribution shows bins 0–7 ≈ 48–50%, which is broadly consistent with this, but the rationale does not explicitly state ‘I am anchoring to a median that falls just after the deadline and allocating accordingly.’ Instead, it vaguely references ‘substantial mass’ without quantifying.
• Percentile consistency check skipped: The forecast does not publish p10, p25, p50, p75, p90 percentiles. If computed from the bin distribution, a p50 (median) near Aug 12–Aug 20 would be consistent with the conservative survival model’s median. However, without an explicit percentile table, it is difficult to verify that the bin-level distribution matches the forecaster’s own stated confidence levels (‘medium confidence’ per the draft, though not published in the submission).

Reconcile: partial_shift

Reviews consumed: 3

The anchoring reviewer flagged that the draft over-anchors on the optimistic rising-hazard model’s mid-July median despite 69 days of post-first-solve silence, suggesting a shift toward late July/early August and more bin_8 mass. The ceiling reviewer flagged two structural constraints I find compelling: (1) the “publishable result” bar already rejected the Ramsey Book Graphs partial progress (n≤56), and (2) Epoch AI’s scaffold testing of GPT-5.4 Pro found zero other solves — implying near-term bins (bin_0, bin_1) should be very small. The math reviewer correctly noted the draft’s bin_0–7 sum (~48%) was below its stated 55–60% inside-view target, indicating an implicit reconciliation drift toward the contrarian view; I accept this as the correct synthesis given the structural evidence. I partially shifted: reduced bins 0–1 toward near-zero (institutional latency + 69-day silence), kept the modal mass in bins 4–5 (mid-July, matching rising-hazard median), held bin_8 large (~0.20, conservative-model median near deadline + publishable-result filter delaying announcement), and increased bin_9–10 mass to reflect harder remaining tiers. Total P(by Aug 12) ≈ 47%, sitting between the conservative survival model (48%) and the contrarian view, which feels right given the ceiling constraints.

Rejected critiques

• [math] Ensemble reconciliation of three variants was not transparently shown; published bins 0–7 (~48%) drifted below stated inside-view 55–60%.

  • Why rejected: I accept the diagnostic but reject the implication that the lower value is wrong. The drift toward ~47% is actually the correct synthesis once the ceiling reviewer’s publishable-result constraint and 69-day empirical silence are integrated — these structural factors justify pulling below the inside-view’s initial 55–60%. So I did not ‘fix’ this by raising bins 0–7 back toward 58%.

• [math] Inside-view stated ‘roughly uniform’ spread but actual distribution is peaked at bins 4–5.

  • Why rejected: The peaked-at-mid-July shape is defensible and aligns with the rising-hazard model’s median; the ‘roughly uniform’ phrasing was loose but the underlying shape is appropriate. No change needed.

Final Distribution (date ranges)

Evidence Sources (40)

1. GPT-5.4 solves its first open math problem from FrontierMath benchmark — GPT-5.4 Pro solved the first FrontierMath Open Problem (a Ramsey-style problem on hypergraphs) in approximately late March 2026, verified by Epoch AI and the original problem author.
2. FrontierMath: Tiers 1-4 - Epoch AI — As of May 11, 2026, Epoch AI is conducting an AI-assisted review of FrontierMath Tiers 1-4 that has flagged fatal errors in about a third of problems; updated scores will only be released after thorough human review is completed.
3. First AI solution on FrontierMath: Open Problems - Epoch AI Substack — Epoch AI officially announced on March 23, 2026 (via Substack) that GPT-5.4 Pro solved the first FrontierMath Open Problem — a Ramsey-style problem on hypergraphs contributed by Will Brian, rated “Moderately Interesting.”
4. GPT-5.4 Pro solved a Ramsey hypergraph problem in FrontierMath — As of the Epoch AI announcement (March 23, 2026), the FrontierMath Open Problems benchmark contains 15 researcher-level problems, with 1 solved (the hypergraph problem). The remaining 14 problems span 4 tiers: Moderately Interesting (3 remaining), Solid Result (5), Major Advance (3), Breakthrough (3).
5. FrontierMath - Wikipedia — Wikipedia (as of May 16, 2026) confirms FrontierMath has 14 bespoke intractable problems remaining unsolved as of that date, with only the hypergraph problem solved. No second solve had been announced by then.
6. FrontierMath: Open Problems - Epoch AI — The FrontierMath Open Problems page snippet as of ~May 2026 shows: “The problem page remains up: see it for more info on an AI-generated solution and subsequent human elaboration. 2026-02-24: We added two problems to the…” — indicating the site had updates but still only one solved problem confirmed.
7. An EpochAI Frontier Math open problem may have been solved - Reddit — A Reddit post from late March 2026 claimed “An EpochAI Frontier Math open problem may have been solved for [the first] time,” discussing GPT-5.4 possibly solving a problem (this appears to be referring to the Ramsey hypergraph problem before the official Epoch AI confirmation).
8. Will AI models solve at least 2 FrontierMath Open Problems before… - Manifold Markets — Manifold Markets shows a prediction that AI models will solve at least 2 FrontierMath Open Problems before August 2026 at ~80% probability, and a related market tracking timing shows 44% probability “Before July 2026” and 50% probability “Before August 2026.”
9. GPT-5.4 set a new record on FrontierMath - Epoch AI Substack — As of Epoch AI’s March 23, 2026 announcement, no second FrontierMath Open Problem had been solved; Epoch’s own evaluation found GPT-5.4 Pro did not solve any Open Problems in their scaffold beyond the hypergraph problem (“It did not solve any problems. It made some novel observations on one problem”).
10. When an AI Writes the Math Paper - EMSI — The solved hypergraph problem was described as possibly “low-hanging fruit” among the open problems set — it took ~250k tokens to solve and was computational combinatorics (constructing specific hypergraphs), closer to computation than traditional pen-and-paper proof.
11. GPT-5.5 was used to flag fatal errors in FrontierMath problems - Reddit — Epoch AI announced that GPT-5.5 was used to flag fatal errors in about a third of FrontierMath (Tiers 1-4) problems, with updated scores to be released. This is a benchmarking quality issue, distinct from FrontierMath Open Problems.
12. Epoch AI Substack - Introducing FrontierMath: Open Problems — The FrontierMath benchmark was introduced in January 2026 with open problems. The first (and only confirmed) solve was announced March 23, 2026 — approximately 2 months after the open problems were launched. No second solve was announced in the ~10 weeks from March 23 to May 31, 2026.
13. DeepMind AI Co-Mathematician hits 48 percent on FrontierMath Tier 4 — Google DeepMind’s AI Co-Mathematician (built on Gemini 3.1 Pro) achieved 48% on FrontierMath Tier 4 (23/48 problems) as of May 7-9, 2026 — significantly better than prior models but this is FrontierMath Tiers 1-4, NOT the Open Problems benchmark. The DeepMind system helped Oxford mathematician Marc Lackenby solve an open conjecture from the Kourovka Notebook, but this was not a FrontierMath Open Problem.
14. The Epoch Brief - May 22, 2026 - Epoch AI Substack — Epoch AI’s May 22, 2026 brief announced FrontierMath: Open Problems (FM:OP) in-person workshops starting May 26 in New York City, but made no mention of a second FrontierMath Open Problem being solved — strongly suggesting no new solve had been announced as of that date.
15. Explicit Deformations of Algebras - Epoch AI — Epoch AI removed the “Explicit Deformations of Algebras” problem from the FrontierMath Open Problems benchmark on 2026-03-12, after GPT-5.2 Pro solved it, determining the solution did not meet the publishable-result bar. This was removed BEFORE the hypergraph solve was announced, meaning it does not count as a qualifying solve for the question.
16. FrontierMath: Open Problems - Epoch AI — The current FrontierMath Open Problems page (as of late May 2026) shows: Moderately interesting 1/4 SOLVED, Solid result 0/5 SOLVED, Major advance 0/3 SOLVED, Breakthrough 0/3 SOLVED. Problems listed as unsolved include: Hadamard Matrices, Ramsey Numbers for Book Graphs, Finiteness Problem for Diophantine Equations, Arithmetic Kakeya Conjecture, Degree vs Sensitivity for Boolean Functions, Surface with High Number of Singularities, Large Steiner Systems, and others.
17. FrontierMath Tier 4 - Epoch AI — The Epoch AI FrontierMath Tiers 1-4 error review (flagging ~1/3 of problems as having fatal errors, announced May 11, 2026) may slow or complicate announcements for the Open Problems subset since it signals Epoch AI is in an active quality review mode, potentially delaying evaluations.
18. Survival forecast calculation — Survival forecast for the next FrontierMath Open Problem announcement: using a hazard rate of ~1 solve per 60 days (the benchmark solved its first problem ~55 days after launch) but with falling hazard (shape=0.7, reflecting the remaining problems being harder), with 69 days already elapsed since the first solve, the model gives P(second solve announced by Aug 12, 2026) ≈ 47.8%. Median remaining wait from today is ~78 days, so median date for second solve is around mid-August 2026 — right at the deadline. P10–P90 remaining wait: ~11–315 days.
19. Workshops - FrontierMath Open Problems | Epoch AI — Epoch AI is running FM:OP workshops in Berkeley, Los Angeles, Toronto, Princeton, NYC, Boston, and London during late May and early June 2026 to DEVELOP NEW PROBLEMS for the benchmark — not to solve existing ones. These workshops are aimed at adding new unsolved problems from working mathematicians.
20. GPT-5.4 set a new record on FrontierMath - Epoch AI Substack — On March 5, 2026 (before the first FrontierMath Open Problems solve), Epoch AI tested GPT-5.4 Pro on FrontierMath Open Problems and reported “It did not solve any problems. It made some novel observations on one problem, but of a form that the author had anticipated and characterized as relatively uninteresting.” This shows even frontier models were not initially solving open problems.
21. Reddit - r/OpenAI - chart showing unsolved math problems solved — As of a recent Reddit post (May 2026), a chart showed AI solving nine novel math problems and 44 new conjectures, with some unsolved for 50 years. The pace of AI mathematical discoveries appears to be accelerating in 2026.
22. FrontierMath: Open Problems - Unsolved Mathematical Challenges | Epoch AI — The FrontierMath Open Problems page explicitly shows only 1 of 15 problems solved (the Ramsey hypergraph problem) as of the current Epoch AI website, confirming no second solve has been officially announced through late May 2026.
23. Are AI benchmarks doomed? - Epoch AI Substack — As of early May 2026, Epoch AI’s benchmarking team (in a May 1 podcast) still describes FrontierMath Open Problems as an active benchmark for “real unsolved math research problems” — no second solve is mentioned, consistent with no second solve having been announced yet.
24. Multiple sources - synthesized — A key contrarian concern: the 69+ days since the first solve (March 23 to May 31, 2026) with zero second announcement suggests the remaining 13 unsolved problems are substantially harder, which undermines a simple “rate continues from first solve” extrapolation.
25. OpenAI claims it solved an 80-year-old math problem - TechCrunch — OpenAI (May 20, 2026) claimed an unreleased reasoning model disproved a famous 78-80 year old geometry conjecture, but this is NOT a FrontierMath Open Problem — it is a separate mathematical breakthrough, distinct from the Epoch AI FrontierMath benchmark.
26. Survival forecast computation — Survival forecast model 1 (rising hazard, shape=1.5, rate=0.067/week = ~1 solve per 15 weeks since first announcement, with 10 weeks elapsed and 10.4 weeks remaining): P(next qualifying solve announced by Aug 12) ≈ 71.6%. Median remaining wait ≈ 6.1 weeks (from May 31), so expected announcement ~mid-July 2026. P10-P90 range = 1-17 more weeks from now.
27. Survival analysis synthesis — Two survival forecast models bracket the probability of a qualifying announcement by Aug 12: an optimistic rising-hazard model gives ~72%, a conservative constant-hazard model gives ~29%. The wide range reflects uncertainty about whether the remaining 14 unsolved problems are much harder or comparable to the first solved problem.
28. FrontierMath: Tiers 1-4 - Epoch AI — The FrontierMath AI-assisted error review (announced 2026-05-11) explicitly applies only to “FrontierMath: Tiers 1-4” — it is NOT mentioned anywhere on the FrontierMath: Open Problems pages, confirming the error review scope is restricted to Tiers 1-4 and does not encompass the Open Problems subset.
29. A Ramsey-style Problem on Hypergraphs | Epoch AI — The Ramsey-style Hypergraphs problem was solved by GPT-5.4 Pro (elicited by Kevin Barreto and Liam Price); subsequently Opus 4.6 (max), Gemini 3.1 Pro, and GPT-5.4 (xhigh) also solved it in a general scaffold, but GPT-5.2 (xhigh), Opus 4.5 (max), and Kimi K2.5 Thinking could not solve it. No second FrontierMath open problem has been solved.
30. Data on AI Capabilities and Benchmarking | Epoch AI — The Epoch AI benchmarks page (epoch.ai/benchmarks, updated May 31, 2026) lists only one FrontierMath Open Problems solve announcement: “Mar. 23, 2026 — AI has solved one of the problems in FrontierMath: Open Problems.” No subsequent second-solve announcement appears in the benchmarking updates through May 31, 2026.
31. FrontierMath: Open Problems - Unsolved Mathematical Challenges (changelog snippet) — The FrontierMath: Open Problems page records its own independent quality action: “2026-03-05: We removed one problem from the benchmark, as we have determined that any solution would not meet our bar of being a publishable result in its own right” — this was the ‘Explicit Deformations of Algebras’ problem. This removal was specific to the Open Problems set and unrelated to the Tiers 1-4 AI-assisted error review.
32. FrontierMath: Open Problems - Unsolved Mathematical Challenges | Epoch AI — The FrontierMath Open Problems page lists exactly 15 problems across four tiers: 4 “Moderately interesting,” 5 “Solid result,” 3 “Major advance,” and 3 “Breakthrough,” spanning fields including Number Theory, Combinatorics, Algebraic Geometry, and Topology/Geometry.
33. Benchmarking AI on unsolved math problems - Epoch AI (Open Problems about page) — The Open Problems benchmark was initially released as a pilot with 14 problems; after adding one more, it reached 15 total. The current count is 15 qualifying problems (4+5+3+3 across tiers), with 1 solved and 14 unsolved, and no AI-assisted error-review notice appears on the Open Problems page or its about page as of the most recent fetch.
34. A Ramsey-style Problem on Hypergraphs | Epoch AI — The only FrontierMath Open Problem solved by AI to date is “A Ramsey-style Problem on Hypergraphs” (rated “Moderately Interesting”), first solved by GPT-5.4 Pro in approximately March-April 2026; GPT-5.4 (xhigh), Gemini 3.1 Pro, and Claude Opus 4.6 (max) were subsequently also able to solve it. Earlier models GPT-5.2, Opus 4.5, and Kimi K2.5 Thinking failed to solve it in Epoch AI’s scaffold testing.
35. Ramsey Numbers for Book Graphs | Epoch AI — The “Ramsey Numbers for Book Graphs” open problem (Moderately Interesting) remains unsolved as a full problem, but as of a May 2, 2026 progress update, AI-assisted researchers (using GPT-5.2 Pro and GPT-5.4 Pro scaffolds combined with Claude Code/Opus 4.6) found constructions for all n ≤ 56, resolving the previous “single challenge” of n=50 and prompting Epoch AI to raise the single challenge to n=100. This constitutes partial progress but not a full solve.
36. Hadamard Matrices | Epoch AI — The “Hadamard Matrices” open problem (find a Hadamard matrix of order 668, rated “Moderately Interesting”) remains unsolved by AI as of the latest Epoch AI page data. The problem page shows AI attempts but the problem status remains “Unsolved.” The problem author estimates 1–4 weeks for an expert human and rates solvability at 95–99%.
37. The Arithmetic Kakeya Conjecture | Epoch AI — The “Arithmetic Kakeya Conjecture” (rated “Solid Result”) remains unsolved by AI. The problem requires constructing a combinatorial object that improves the best-known upper bound AK(γ) where γ ≈ 1.675 (the largest root of x³ − 4x + 2). The problem page notes that this method cannot work for α < 3/2, meaning it cannot fully establish the Kakeya conjecture.
38. GPT-5.5 Ramsey Number Proof: First Lean-Verified LLM Result — GPT-5.5 produced a formally Lean 4-verified result on a sub-problem of classical Ramsey theory (establishing a new lower bound that no published human work had previously reached), as reported April 25, 2026. However, this is distinct from — and does not constitute a solve of — the “Ramsey Numbers for Book Graphs” FrontierMath Open Problem, which remains officially unsolved on Epoch AI’s tracker.
39. Fields Medalist says ChatGPT 5.5 Pro delivered “PhD-level” math research in under two hours with zero human help — Fields Medal-winning mathematician Timothy Gowers reported in approximately late May/June 2026 that GPT-5.5 Pro solved open mathematical problems from additive number theory (from a paper by Mel Nathanson) in about an hour at “PhD-thesis level,” warning “We will face a crisis very soon.” These were open problems in the literature, not the specific FrontierMath Open Problems benchmark.
40. Which Epoch AI FrontierMath open problem will be solved next? | Manifold — Manifold prediction market (as of ~April 23, 2026) placed roughly equal 26% probability on each of Hadamard Matrices, Ramsey Numbers for Book Graphs, and Finiteness Problem for Diophantine Equations being the next FrontierMath Open Problem to be solved, with 13% on a “Solid result” tier problem and 9% on a “Major advance” or “Breakthrough” problem.

wooley_rhino • run 96dfbf54 • $3.23 • 6m18s

Extracted Date Percentile Values: {0.1: 1780272000.0, 1.0: 1780617600.0, 5.0: 1782172800.0, 10.0: 1783641600.0, 15.0: 1784505600.0, 20.0: 1785196800.0, 25.0: 1785888000.0, 30.0: 1786406400.0, 35.0: 1787011200.0, 40.0: 1788566400.0, 45.0: 1790294400.0, 46.0: 1790726400.0, 47.0: 1791158400.0, 48.0: 1791763200.0, 49.0: 1792454400.0, 50.0: 1793145600.0, 51.0: 1794096000.0, 52.0: 1794960000.0, 53.0: 1795824000.0, 54.0: 1796860800.0, 55.0: 1797897600.0, 60.0: 1802649600.0, 65.0: 1809129600.0, 70.0: 1818288000.0, 75.0: 1827619200.0, 80.0: 1840752000.0, 85.0: 1856649600.0, 90.0: 1874966400.0, 95.0: 1922313600.0, 99.0: 2051222400.0, 99.9: 2366841600.0}

Model’s Answer:

Forecasting Analysis: Next FrontierMath Problem Solve

Understanding the Question

This question asks for the date when Epoch AI will publicly announce that an AI has solved another qualifying FrontierMath open problem (beyond the already-solved Ramsey-style hypergraph problem). Resolution depends entirely on Epoch AI’s announcement timing, not when the actual solve occurs.

Key Evidence Assessment

Current Status (as of May 31, 2026):

• One solve to date: Ramsey-style hypergraph problem by GPT-5.4 Pro (announced March 23, 2026)
• Remaining unsolved: 3 Moderately Interesting, 5 Solid Result, 3 Major Advance, 3 Breakthrough problems
• GPT-5.4 Pro “did not solve any problems” beyond the hypergraph case despite setting records on easier tiers
• No changelog updates since March 5, 2026; no indication of imminent solves
• Workshops scheduled for late May/early June 2026 (problem development, not solving)

Time Analysis:

• ~69 days since last solve announcement
• ~73 days until resolution deadline (August 12, 2026)
• We’re already as far from the last solve as the entire remaining window

Critical Assumptions to Question

Assumption 1: Current AI capabilities are sufficient for additional solves

• Evidence against: GPT-5.4 Pro explicitly failed to solve any additional open problems
• Probability of failure: ~60% (major capability jump likely needed)
• Impact: If correct, dramatically reduces probability of near-term solve

Assumption 2: New model releases or breakthroughs will occur within window

• Timeline for major releases is typically 6-12+ months between frontier models
• No public announcements of imminent releases
• Probability this assumption is wrong: ~70%

Assumption 3: Problem difficulty is well-calibrated for near-term solving

• The easiest category still has 3 unsolved problems despite best efforts
• Problems may be harder than the single solved one
• Probability this assumption is wrong: ~50%

Probability Estimate

Given:

• Only 1 solve in ~4-5 months of benchmark operation
• Current SOTA (GPT-5.4 Pro) explicitly unable to solve remaining problems
• No indicators of imminent progress
• Remaining problems likely harder than the solved one

P(solve announced by Aug 12, 2026): ~28-32%

This is lower than a simple “1 solve per 5 months = maybe another in 2.5 months” extrapolation because:

• Low-hanging fruit already picked
• Explicit evidence current systems cannot solve remaining problems
• No indication of capability jumps in the pipeline

Why Earlier Outcome Could Occur

• Unpublicized work may already be close to a solve
• New model release (GPT-5.5 or competitor) could unlock capability jump
• Workshops may identify easier qualifying problems
• Continued intensive effort on remaining Moderately Interesting problems could succeed
• ~28-30% probability mass within window

Why Later Outcome Could Result

• GPT-5.4 Pro explicitly failed on all remaining problems
• No changelog activity in nearly 3 months
• More difficult problems require fundamental capability improvements
• Major model releases typically take 6-12+ months
• If no solve by Aug 12, could take months or years for next breakthrough
• ~70% probability mass after August 12

Distribution Construction

Within-window distribution (30%):

• Early June: Very unlikely, no signs of imminent announcement (~2%)
• Mid-late June: Possible if work in progress (~4%)
• July: Reasonable if new developments occur (~10%)
• August 1-12: Accumulated probability from ongoing efforts (~14%)

After-window distribution (70%):

• Late 2026: Continued progress possible (~20%)
• 2027: More capability development time (~25%)
• 2028+: Harder problems requiring major breakthroughs (~25%)

Final Rationale

The single solve in March 2026 demonstrated feasibility but also revealed limitations—the same model failed on all remaining problems. With no evidence of imminent capability jumps, the most likely scenario is no solve within the remaining 73 days. However, AI progress is unpredictable, and intensive focused work could yield surprises. I’m setting a ~30% probability for resolution within the window, heavily weighted toward late July and early August, with 70% probability mass extending into future months and years with decreasing density.

Percentile .1: 2026-06-01 Percentile 1: 2026-06-05 Percentile 5: 2026-06-23 Percentile 10: 2026-07-10 Percentile 15: 2026-07-20 Percentile 20: 2026-07-28 Percentile 25: 2026-08-05 Percentile 30: 2026-08-11 Percentile 35: 2026-08-18 Percentile 40: 2026-09-05 Percentile 45: 2026-09-25 Percentile 46: 2026-09-30 Percentile 47: 2026-10-05 Percentile 48: 2026-10-12 Percentile 49: 2026-10-20 Percentile 50: 2026-10-28 Percentile 51: 2026-11-08 Percentile 52: 2026-11-18 Percentile 53: 2026-11-28 Percentile 54: 2026-12-10 Percentile 55: 2026-12-22 Percentile 60: 2027-02-15 Percentile 65: 2027-05-01 Percentile 70: 2027-08-15 Percentile 75: 2027-12-01 Percentile 80: 2028-05-01 Percentile 85: 2028-11-01 Percentile 90: 2029-06-01 Percentile 95: 2030-12-01 Percentile 99: 2035-01-01 Percentile 99.9: 2045-01-01

The platform’s small curated problem set and strict public-verification rule limit the number of qualifying announcements before the deadline.

Publishability review, verifier-access control, and contributor write-up requirements are durable frictions that lengthen timelines.

Recent capability improvements yielded one public solution and an additional near-miss that demonstrates attainable progress on specific problems.

Several constructive-search targets carry concentrated upside because authors report short expert timelines and high solvability for particular instances.

A privately vetted candidate passing the platform’s review or the addition of an easily verifiable problem would be the main inflection events.

Conversely, qualification failures, protracted write-up reviews, or delayed disclosure will push plausible solutions past the cutoff.

Major uncertainties are opaque private runs, the concentration of verifier access among a few actors, and the platform’s discretionary publishability judgments.

Sensitivity is high: one undisclosed verified solution materially raises resolution odds, while shared bottlenecks compress independent paths and limit near-term multiplicity.

Forecast: 2026-07-13 (median; 1784421545.8734 UTC epoch seconds).

• Big bottleneck is verification/public announcement. Only one organization currently holds the verifier-access role, so even if progress happens, the public “confirmed solve” can lag.
• There’s evidence that solves can come in bursts, but not reliably. After a previous confirmed solve in late March, there’s been a long quiet stretch with no new confirmation yet.
• External market odds tilt toward mid-summer, not late fall. Signals suggest a meaningful chance of a solve by early July, with the odds still rising through early fall.
• A “Moderately Interesting” subset is viewed as more reachable soon. If one of those problems is truly within current capability, it could trigger a relatively near-term announcement.
• There’s a heavy tail beyond the deadline. Even though mid-summer is the center, a substantial fraction of outcomes are still delayed well past mid-August.