AI model scores ≥ 90% on FrontierMath Benchmark before 2027?

Resolution Criteria

This market will resolve to "Yes" if a state-of-the-art (SOTA) AI model achieves a score of 90% or greater on the FrontierMath Exam by December 31, 2026, 11:59 PM ET. Otherwise, the market will resolve to "No". The primary resolution source will be information from EpochAI however a consensus of credible reporting may also be used.

Prediction markets currently show relatively low confidence that any AI model will score 90% or higher on the FrontierMath benchmark before the end of 2026, with volume concentrated on the 'No' outcome. FrontierMath is widely regarded as one of the most demanding mathematical reasoning benchmarks available, and the market reflects the substantial gap between current SOTA performance and the 90% threshold. Resolution depends on verified results published or confirmed by EpochAI by 31 December 2026.

Market structure

This is a binary yes/no market with a single resolution question. Volume is heavily concentrated on the 'No' outcome, indicating that the majority of market participants do not anticipate a sub-two-year breakthrough to the 90% threshold. The primary resolution source is EpochAI, with credible consensus reporting as a fallback. The deadline is 31 December 2026 at 11:59 PM ET, with no extension mechanism described.

Background

FrontierMath is a benchmark developed to test advanced mathematical reasoning in AI systems, featuring problems that require graduate-level and research-level mathematical knowledge across areas such as number theory, algebraic geometry, and combinatorics. It was designed explicitly to resist near-term saturation, with problems contributed by professional mathematicians to ensure they cannot be solved by pattern-matching against training data alone. When the benchmark was introduced, leading AI models scored in the low single digits, making a 90% threshold a dramatic step change. The benchmark has attracted significant attention in AI research circles as a credible signal of genuine mathematical reasoning capability, distinct from more easily gamed standardised tests. Progress on reasoning benchmarks has accelerated with the development of models using chain-of-thought and extended inference techniques.

Key factors

The pace of improvement in AI mathematical reasoning is the central variable. Models using extended test-time compute and reinforcement learning from verifiable rewards have shown rapid gains on structured reasoning tasks, which could narrow the gap faster than historical rates suggest. However, FrontierMath problems are deliberately novel and not reproducible from prior literature, which limits the benefit of scale alone. A key dependency is whether EpochAI or the benchmark's maintainers formally evaluate new models against the full benchmark, since unofficial or partial evaluations may not satisfy resolution criteria. The accessibility of the benchmark to leading AI laboratories — whether open or subject to controlled evaluation — affects how quickly results can be generated and verified. Any significant architectural shift, such as the integration of formal proof systems or enhanced symbolic reasoning, could create a discontinuous jump in scores. Conversely, if leading laboratories deprioritise benchmark performance in favour of applied capability, measured progress may stall.

FAQ