Atlas

Benchmarks

PatchBench (OpenAI internal)

OpenAI's internal benchmark of repairing security vulnerabilities with patches that preserve intended behavior, reported in the GPT-6.2 launch post. Task count and grading are not published. Diagnostic; OpenAI runs only.

Models
4
Top score
79.7
Leader
GPT-6.2 Astra
Updated
Nov 17, 2026
Resolution
—
Saturation
—

No signal or difficulty estimate. Held out of the fit by design.

Organizations
4 dated models · 3 frontier modelsClosed weightsOpen weights
Data: OpenAIatlas.village.dev

PatchBench (OpenAI internal) leaderboard

Loading Atlas data…