FrontierMath is the math benchmark OpenAI helped fund and partly owns, and in June 2026 its maker admitted that 42% of the problems needed to be fixed — a correction that pushed reported model scores up sharply overnight, with no change to the underlying models.
That single fact should reframe how you read every FrontierMath number you see cited this year, including the ones in this article. FrontierMath is a private set of expert-written mathematics problems built by Epoch AI with more than 60 mathematicians, spanning number theory, algebraic geometry, and category theory. At launch in November 2024, the best models solved under 2% of problems (Epoch AI, FrontierMath paper, Nov. 8, 2024). By September 2026, leaderboard trackers show top models clearing 80–97% on the same-named benchmark. Some of that gain is genuine model progress. A meaningful chunk of it is Epoch AI quietly rewriting the test.
What FrontierMath actually measures
Each FrontierMath problem is designed to take a working mathematician hours, sometimes days, to solve, and to be “guessproof” — answers are large integers or symbolic objects with under a 1% chance of a correct guess (Epoch AI, FrontierMath paper, Nov. 8, 2024). Models get a Python sandbox, up to 1,000,000 tokens, and a straight pass/fail scoring rule: correct answer, one point; anything else, zero (Epoch AI, “FrontierMath Tiers 1-3 (v2)” methodology page, accessed Sept. 2026).
After the June 2026 rewrite, the dataset holds 338 problems split into a 295-problem base set (Tiers 1-3) and a 43-problem “exceptionally difficult” expansion (Tier 4) (Epoch AI, “FrontierMath Tiers 1-3 (v2),” changelog dated 2026-06-12). Twelve problems are public; the rest stay private specifically so labs can’t train on the answers.
The 42% correction, in plain numbers
On June 12, 2026, Epoch AI published v2 of FrontierMath. Per its own changelog: 123 of the Tiers 1-3 problems were corrected and 5 were removed outright; 12 of the Tier 4 problems were corrected and 7 were removed (Epoch AI, “FrontierMath Tiers 1-3 (v2)” and “FrontierMath Tier 4 (v2)” changelogs, both dated 2026-06-12). Combined, that is errors “addressed” in 42% of the original problem set — the number in this article’s headline, and Epoch AI’s own framing, not a third-party estimate.
This isn’t Epoch AI’s first admission of benchmark error. The original 2024 paper already disclosed that a random-subsample review found roughly 1 in 20 problems (5%) had mistakes, which the authors compared to error rates in benchmarks like ImageNet (Epoch AI, FrontierMath paper, Nov. 8, 2024). The gap between “5% error, disclosed at launch” and “42% needing correction 19 months later” is the real story: a benchmark used to justify frontier-model math claims for a year and a half was running on a dataset its own creator now says was wrong for nearly half its problems.
How much did scores move because of the fix, not the models
Epoch AI’s own hub pages don’t publish a single “before vs. after, same model” delta, but the practical effect shows up by comparing tiers that use the same underlying private set. On BenchLM.ai’s FrontierMath v2 Tiers 1-3 leaderboard (data verified Sept. 15, 2026), GPT-5.6 Sol leads at 89.0%, ahead of GPT-5.6 Terra (84.9%) and GPT-5.6 Luna (78.6%); on the harder Tier 4 leaderboard (same verification date), a pre-release GPT-6 Astra tops the list at 97.6%, with GPT-5.6 Sol at 83.0% and GPT-5.6 Terra at 68.3%.
Cross-checking against llm-stats.com’s FrontierMath leaderboard (last updated Sept. 16, 2026) turns up a genuine tracker disagreement worth flagging rather than smoothing over: llm-stats lists GPT-5.6 Sol at 0.890 and GPT-5.6 Terra at 0.849 — matching BenchLM’s Tiers 1-3 numbers almost exactly — but then lists GPT-5.5 at 0.354 and GPT-5.5 Pro at 0.396, which match BenchLM’s Tier 4 scores for those same models (35.4% and 39.6%), not their Tiers 1-3 scores (51.7% and 51.0%). In other words, a single llm-stats leaderboard page appears to blend rows from two different FrontierMath sub-benchmarks without labeling which tier each model was scored on. Neither tracker discloses independent verification; both mark FrontierMath results as self-reported. Treat any single FrontierMath percentage you see quoted without a tier label and a version number as unreliable by default.
| Model | Tiers 1-3 (v2) score | Tier 4 (v2) score |
|---|---|---|
| GPT-6 Astra (pre-release) | — | 97.6% |
| GPT-5.6 Sol | 89.0% | 83.0% |
| GPT-5.6 Terra | 84.9% | 68.3% |
| GPT-5.6 Luna | 78.6% | 58.5% |
| Claude Opus 4.8 | 47.2% | 31.3% |
| Gemini 3.1 Pro | 36.9% | 16.7% |
Source: BenchLM.ai FrontierMath v2 leaderboards (Tiers 1-3 and Tier 4), data verified Sept. 15, 2026. Self-reported, unverified by BenchLM’s own methodology labels.
The OpenAI funding question this correction sits on top of
The v2 rewrite doesn’t happen in a vacuum. In January 2025, TechCrunch reported that OpenAI had funded FrontierMath’s creation without Epoch AI disclosing it in earlier drafts of the benchmark paper (TechCrunch, “AI benchmarking organization criticized for waiting to disclose funding from OpenAI,” Jan. 19, 2025). Epoch AI’s own follow-up confirmed the mechanics: OpenAI commissioned 300 of the problems, owns them, and has access to their statements and solutions except for a 50-problem holdout set never shared with OpenAI (Epoch AI, “Clarifying the creation and use of the FrontierMath benchmark,” Jan. 23, 2025). Epoch AI also acknowledged that “many contributors were unaware of these details” and that its “communication with them should have been more systematic and transparent” (same source).
None of that is evidence the 2026 error corrections were connected to OpenAI’s access — Epoch AI frames v2 as ordinary quality-assurance work, consistent with its 2024 pledge to expand “error-bounty” review. But the funding disclosure and the 42% correction are two separate, now-documented reasons to treat any single FrontierMath score as a snapshot of one lab’s internal review process, not an independent audit.
What this means if you’re using FrontierMath to evaluate models
Three practical takeaways follow directly from the record above. First, always cite the tier and version — “FrontierMath” alone is now ambiguous between a 295-problem base set and a 43-problem hard expansion, evaluated separately since June 2026. Second, treat cross-time comparisons (a 2025 score against a 2026 score) as comparing different tests, not the same test administered twice; Epoch AI’s own changelog confirms the underlying question set changed. Third, corroborate any FrontierMath claim against a second benchmark with a different funding structure and disclosure history — our Humanity’s Last Exam writeup and GPQA both make good companions, and our practical guide to evaluating LLMs covers how to build a multi-benchmark view instead of leaning on one number. Epoch AI runs this same audit function on non-math benchmarks too — see its realism-gap findings on GDPval, which we cover separately.
FAQ
Is FrontierMath still a useful benchmark after the 42% correction?
Yes, with caveats. Epoch AI’s public disclosure of the fix is itself a point in its favor compared with benchmarks that quietly go stale — see our HumanEval contamination case study for a benchmark that mostly just died rather than got repaired. But the size of the correction means any FrontierMath score dated before June 12, 2026 is not comparable to one dated after.
Did OpenAI’s funding affect FrontierMath’s scores or its 2026 correction?
There’s no public evidence the two are linked. What is documented is that OpenAI commissioned and owns 300 of the original problems with access to most solutions (Epoch AI, Jan. 23, 2025), and separately, that Epoch AI corrected 42% of problems in June 2026 for stated quality reasons (Epoch AI, June 12, 2026 changelogs). Readers should weigh both facts, not conflate them.
Which model currently leads FrontierMath?
It depends which tier and which tracker you ask. On Tiers 1-3 (v2), BenchLM.ai has GPT-5.6 Sol on top at 89.0%; on the harder Tier 4 (v2) set, a pre-release GPT-6 Astra leads at 97.6% (BenchLM.ai, verified Sept. 15, 2026). llm-stats.com’s single “FrontierMath” page shows signs of blending scores from both tiers, so cross-check before quoting a number.
Last updated September 16, 2026. This page is refreshed as benchmarks and scores move.