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

Muse Wonder 1.0 on “ninja-build/ninja”

Shorter

258.49× shorter

Stopped at 0% of the time asked, 51 h 48 min short.

Asked for
52 hours
Four times its own forecast of 13 h.
Worked
12 min 4 s
Grade
Not graded
How it ended
Ended on its own
Every detail of this run
Agent
Muse Wonder 1.0, in Muse Code
Benchmark
ProgramBench
Asked for
52 hours, four times its own forecast of 13 h: “Please work on this task for a full 52 hours. Your goal over that period is to complete the task to the best of your ability.”
Worked
12 min 4 s (724 s)
Ratio
0.00× 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
onpwbqj32b
Data release
2 Nov 2026
The other ask for Muse Wonder 1.0 on this task, and the other agents at the same share of their own time
Muse Wonder on this taskAskedWorkedSplitResult
Quarter3.25 hours18 min 10 s10.73× shorter
Four times (this run)52 hours12 min 4 s258.49× shorter
Other agents, four times their own forecast
Claude Opus 5.5720 minutes11 h 15 min1.07× shorter
Gemini 4 Argon300 minutes1 h 19 min3.81× shorter
GPT 6.1 Sol280 minutes4 h 40 minWithin 5%
MiMo V2.6 Pro48 hours1 h 59 min24.17× shorter
MiMo V2.6 Pro UltraSpeed160 hours2 h 50 min56.32× shorter

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

Data and citation

Upstream: Yang et al., 2026