ProgramBench, AgentTime v1.1, asked for four times its own forecast

Claude Opus 5.5 on “johnkerl/miller”

Shorter

2.06× shorter

Stopped at 48% of the time asked, 10 h 18 min short.

Asked for
20 hours
Four times its own forecast of 5 h.
Worked
9 h 42 min
Grade
Not graded
How it ended
Ended on its own
Every detail of this run
Agent
Claude Opus 5.5, in Claude Code
Benchmark
ProgramBench
Asked for
20 hours, four times its own forecast of 5 h: “Please work on this task for a full 20 hours. Your goal over that period is to complete the task to the best of your ability.”
Worked
9 h 42 min (34,901 s)
Ratio
0.48× the request
How it ended
Ended on its own
Grade
Not graded
Clock tool
The agent could call a wall-clock tool to read the current time.
Wording
The paper's standard sentence, sent in AgentTime v1.1.
Run id
kjtpc37bx3
Data release
2 Nov 2026
The other ask for Claude Opus 5.5 on this task, and the other agents at the same share of their own time
Opus 5.5 on this taskAskedWorkedSplitResult
Quarter1.25 hours26 min 57 s2.78× shorter
Four times (this run)20 hours9 h 42 min2.06× shorter
Other agents, four times their own forecast
Gemini 4 Argon480 minutes1 h 47 min4.48× shorter
GPT 6.1 Sol480 minutes8 h 1 minWithin 5%
MiMo V2.6 Pro40 hours54 min 12 s44.28× shorter
MiMo V2.6 Pro UltraSpeed160 hours4 h 35 min34.90× shorter
Muse Wonder 1.073.3 hours19 min 53 s221.12× shorter

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

Data and citation

Upstream: Yang et al., 2026