AI in Mathematics: capabilities and evidence

What has AI demonstrated?

1.0 mean score in the official export

Test

gpt-6.1-sol_max; FrontierMath Tier 4 v2; run started 29 September 2026.

Comparison

Machine-checkable answer reference; per-run denominator absent from CSV.

System and version

FrontierMath Tier 4 v2 / official CSV · 2026-09-29

Evidence checked

2026-10-20

Corrected June 2026 benchmark; 43 Tier-4 problems, two public, evaluations normally private. Exact per-run denominator and full logs absent in CSV; numeric answers are not research-proof verification.

What does the field require?

Solve a specified difficult mathematical problem

• Understand a formalizable problem

  FrontierMath supplies the problem; run-level details remain incomplete.

  FrontierMath Tier 4 v2 / official CSV
• Find a correct answer

  Current Tier-4 results show strong answer finding under tool and token budgets.

  FrontierMath Tier 4 v2 / official CSV
• Supply a complete verified proof

  Numerical answer grading does not test proof completeness.

  FrontierMath Tier 4 v2 / official CSV

Create a new mathematical research direction

• Choose useful conjectures

  Released manuscripts do not supply a representative discovery denominator.

  OpenAI mathematics release, updated October 15
• Develop a sustained argument

  Selected research outputs support partial capability, not general autonomy.

  OpenAI mathematics release, updated October 15
• Validate novelty and proof

  No matched broad autonomous novelty and proof audit.

  OpenAI mathematics release, updated October 15

Research tasks

Can AI replace the whole workflow?

The results above measure specific tasks. The field assessment also requires completion of the connected workflow, including the requirements listed here. A field-wide replacement percentage has not been established.

Sources and related evidence

• FrontierMath Tier 4 v2 / official CSV
• OpenAI mathematics release, updated October 15

Compare related fields