Is FrontierMath Tiers 1–3 still meaningful?

Known problems

• OpenAI funds it and can see part of it.^{[2][2] GPT-6 Astra benchmarks explained Vellum Secondary}
• Version 2 fixed 123 Tier 1–3 problems. Epoch's ceiling estimate for the original set was only 57%.^{[3][3] FrontierMath Tiers 1-3 v2 Epoch AI Primary[1][1] Benchmark data bundle Epoch AI Primary}

What to look at instead

• FrontierMath Tier 4 Status:Saturated No. A model solved every problem three months after version 2 launched.
• FrontierMath Erdős Status:Still useful Yes. It still leaves almost all of its room at the top, though the set is small.

Where it appears in the report

Chapter 1, Benchmarks

• How benchmarks get replaced
• Table 3. How each area replaced its saturated benchmarks, oldest first in “Go deeper: Lineage of replacements, by area”
• Table 5. Reasoning and knowledge benchmarks: what they measure, who maintains them, the top score on 5 November 2026 and the known flaw in “Go deeper: Reasoning and knowledge benchmarks (17)”