LuminaBench
Reasoning benchmark / Epoch AI

FrontierMath Tiers 1–3

Lumina VerifiedLumina checked the published sources, versions, model labels, score extraction and conversion. External evaluations were run by their publishers; Lumina has not independently rerun or reproduced them.

Research-level mathematics across Tiers 1–3, using the private v2 evaluation set.

Version 2 · PrivateVerified 2 November 2026122 published resultsOfficial source

FrontierMath Tiers 1–3 · Lumina Score

Epoch AI · v2 · 285 private problems

LuminaBench
Show
20 of 89 models · best variant per model
03006009001,2001,115GPT-6.1 SolMax1,115GPT-6 AstraMax1,108Claude Opus 5.5Max1,107GPT-6.1 TerraMax1,105Claude Fable 5.1Max1,104GPT-6 SolMax1,102GPT-5.6 SolMax1,101Claude Sonnet 5.5Max1,098GPT-5.5 ProxHigh1,097Claude Fable 5Max1,094GPT-5.6 TerraMax1,094Gemini 4 ArgonHigh1,093Claude Opus 5Max1,092GPT-5.5xHigh1,084GPT-5.4 ProxHigh1,083Muse WonderMax1,083GPT-5.6 LunaMax1,077Claude Opus 4.8Max1,074GPT-6 LunaMax1,073GPT-5.4xHigh

Epoch AI published accuracy and standard error; unspecified effort remains unknown. Missing metrics are omitted from their graph; quality scores remain available.

Results table

Source & conversion

Epoch AI’s corrected FrontierMath v2 evaluation on the private Tiers 1–3 set.

Proportion of correctly solved private problems, displayed as a percentage. Exported standard errors are retained; no Python runtime, generated tokens, costs or tool-call measurements are published in this CSV.

How the Lumina conversion works

1,000 + 75 × ((original score − 51.7772152638828) / 27.381433968492303). The result is rounded once. Scores may tie.

Frozen initial eligible Epoch cohort. Each exact model release contributes one reference value: the arithmetic mean of all its retained observed configuration scores, so releases with more settings do not receive extra calibration weight. This is normalization only, not a combined index or an estimate. Subsequent source refreshes never refit this calibration.

Calibration
frontiermath-tiers-1-3-v2-z75-v1 · experimental
Reference mean
51.7772152638828
Population SD
27.381433968492303
Coverage
Standalone Reasoning benchmark.
Identity, exclusions & attribution

Exact model releases and explicit native effort settings are retained. Unknown effort remains unknown; labelled token budgets are a separate setting axis.

Dataset size is published; per-row valid sample counts, trial counts, serving API provider and unreported harness versions remain unknown.

0 historical records have unverified current-protocol provenance and remain in the retained original CSV.

Official benchmark · Data and attribution · CC BY 4.0

Publisher terms. Scores and resource figures are attributed to the original publisher.