Data unavailable
The benchmark data could not be loaded.
Loading class relationships…
✨ How it works
The benchmark contains 50 complexity classes and all 2,500 ordered pairs of those classes. Each pair asks whether A ⊆ B: whether every decision problem in A also belongs to B. The reverse direction is a separate question.
The baseline records 1,176 relations known by September 1, 2026. The remaining 1,324 questions were open as of September 1, 2026. A resolution can prove containment, prove noncontainment, or establish that the statement is independent of ZFC.
Accepted proofs are combined with the baseline using the benchmark’s fixed implication rules. Every newly resolved open pair earns one point, including pairs settled as consequences.
For example, an AI proof of P ⊊ NP alone would earn 199 points (a score of 15.03%), as this would also imply P ⊊ PSPACE and settle 197 other open questions.
Score = (newly resolved pairs / 1,324) × 100%.
Leaderboard
Score over time
Frontier models by release date. OAI Internal Model uses the math release date.
| Rank | Model | Score |
|---|---|---|
| 1 | OAI Internal ModelBPL ⊆ L; ∃R ⊆ CH · 5 pairsMath release: | 0.38% |
| 2 | GPT-6 Astra∃R ⊆ CH · 1 pairReleased: | 0.08% |
| 3 | Fable 5.1Released: | ≥ 0.00% |
Class relationships
Click any two classes to compare them, in either order.
✨ Submit a proof
Write the new argument in a Lean project with multiple files, using trusted textbook results. Additional results published by the cutoff need a review of the cited statement and its match to the benchmark’s model; the cited results serve as premises.
Uniform integer-array interfaces cover complete arrays and implicit arrays accessed by index. Materializing an implicit array requires a polynomial bound on its number of entries.
✨ Scope and sources
The questions concern classes of total decision languages, with fixed definitions and circuit-uniformity conventions. Results about promise problems, oracle-relative separations, search problems, or a particular algorithm’s speed require an argument connecting them to those class definitions to resolve a benchmark pair.
11 additional classes remain available inside proofs and implication chains. Points are awarded to pairs of the 50 scored classes.
New containment and noncontainment claims require Lean proof checking and mathematical review. Independence claims require a separate expert review. Existing cited results serve as premises. Independent discovery and prior-exposure checks require separate evidence about a model’s run.
Evaluation details · Cutoff audit · Coverage notes
References
—