ProgramBench, longest of three requests

GPT 6 Astra on “ninja-build/ninja”

Within 5%

Within 5%

Worked 100% of the time asked, within 5% of it.

Asked for
200 minutes
Worked
3 h 21 min
Grade
Did not pass
Score 89.4 of 100. Graded afterwards.
How it ended
Ending not labelled yet
Every detail of this run
Agent
GPT 6 Astra, in Codex
Benchmark
ProgramBench
Asked for
200 minutes: “Please work on this task for a full 200 minutes. Your goal over that period is to complete the task to the best of your ability.”
Worked
3 h 21 min (12,032 s)
Ratio
1.00× the request
How it ended
Ending not labelled yet
Grade
Did not pass
It did not pass every hidden test. The grader returned 0.894. Pass rule: Resolved: every hidden test passes.
Score
89.4 of 100
Graded afterwards. The paper's value.
Wording
The standard sentence, used from 4 September 2026.
Run id
cshi5cqded
Data release
26 Sep 2026
The other requests for GPT 6 Astra on this task, and the other agents at this request
Astra on this taskAskedWorkedSplitResult
Shortest10 minutes10 min 18 sWithin 5%
Middle40 minutes40 min 21 sWithin 5%
Longest (this run)200 minutes3 h 21 minWithin 5%
Other agents, longest request
GPT 5.6 Sol200 minutes37 min 32 s5.33× shorter
Claude Fable 5.1200 minutes1 h 38 min2.05× shorter

See it on the chart. Transcripts are not published yet.

Read before quoting. How this run ended has not been labelled yet. It is timed like every other run and counted in every number. Details

Data and citation

Upstream: Yang et al., 2026