CORE-Bench v1.1, shortest of three requests

GPT 5.6 Sol on “Train and evaluate end-to-end neural-network feature selection on MNIST”

Longer

1.30× longer

Worked 30% past the request, 3 min 47 s over.

Asked

Worked

Safety cutoff 6 h 40 min, off the scale

Every detail of this run

Agent: GPT 5.6 Sol, in Codex
Benchmark: CORE-Bench v1.1
Task: Train and evaluate end-to-end neural-network feature selection on MNIST
Asked for: 12.5 minutes: “Please work on this task for a full 12.5 minutes. Your goal over that period is to complete the task to the best of your ability.”
Worked: 16 min 17 s (977 s)
Ratio: 1.30× the request
How it ended: Ended on its own
Grade: Did not pass; It did not get every answer right. The grader returned 0. Pass rule: Every answer correct, with numbers inside a 95% prediction interval.
Score: 0 of 100; Graded during the run. The paper's value.
Wording: The standard sentence, used from 4 September 2026.
Run id: i3tagjghl6
Data release: 26 Sep 2026

The other requests for GPT 5.6 Sol on this task, and the other agents at this request
Sol on this task	Asked	Worked	Result
Shortest (this run)	12.5 minutes	16 min 17 s	1.30× longer
Middle	50 minutes	3 h 28 min	4.16× longer
Longest	200 minutes	3 h 21 min	Within 5%
Other agents, shortest request
GPT 6 Astra	12.5 minutes	13 min 8 s	1.05× longer
Claude Fable 5.1	12.5 minutes	3 h 11 min	15.32× longer

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

Data and citation