WildClawBench, shortest of three requests
GPT 5.6 Sol on “Triage support messages, investigate hidden context, route issues, and draft replies”
Longer
7.50× longerWorked 650% past the request, 6 min 30 s over.
- Asked for
- 1 minute
- Worked
- 7 min 30 s
- Grade
- Score 87.5 of 100
- Graded during the run.
- How it ended
- Ended on its own
Every detail of this run
- Agent
- GPT 5.6 Sol, in Codex
- Benchmark
- WildClawBench
- Asked for
- 1 minute: “Please work on this task for a full 1 minute. Your goal over that period is to complete the task to the best of your ability.”
- Worked
- 7 min 30 s (450 s)
- Ratio
- 7.50× the request
- How it ended
- Ended on its own
- Grade
- 0.875
- Task grader score, 0 to 1 (a model judge only where the task calls for one). No official pass rule.
- Score
- 87.5 of 100
- Graded during the run. The paper's value.
- Wording
- The standard sentence, used from 4 September 2026.
- Run id
- mcion7x6za
- Data release
- 26 Sep 2026
Related runs
| Sol on this task | Asked | Worked | Split | Result |
|---|---|---|---|---|
| Shortest (this run) | 1 minute | 7 min 30 s | 7.50× longer | |
| Middle | 4 minutes | 5 min 16 s | 1.32× longer | |
| Longest | 15 minutes | 16 min | 1.07× longer | |
| Other agents, shortest request | ||||
| GPT 6 Astra | 1 minute | 3 min 7 s | 3.12× longer | |
| Claude Fable 5.1 | 1 minute | 10 min 3 s | 10.05× longer | |
See it on the chart. Transcripts are not published yet.
Data and citation
Upstream: Ding et al., 2026