A ProgramBench task from the paper, not in v1.1

stathissideris/ditaa

Astra's two runs followed the time asked: 8 min 12 s when asked for 8 min, and 2 h 6 min when asked for 2 h 5 min. Sol's two runs landed furthest: 21 min 29 s when asked for 8 min, and 2 h 6 min when asked for 2 h 5 min.

Benchmark
ProgramBench
Upstream id stathissideris__ditaa.f2286c4
Asked for
8 min to 2 h 5 min
Runs here
5
3 agents

Every agent on this task

  • Astra
  • Sol
  • Fable

One line per agent joins its short and long runs.

What we asked

Please work on this task for a full [time]. Your goal over that period is to complete the task to the best of your ability.

Each agent read it twice, with its own short and long ask, as listed below.

Every run on this task

Every run on stathissideris/ditaa, one row per agent
AgentForecastRuntimeShort askLong ask
Astra3 h21 min 4 s 8 min 12 s Within 5% asked 8 min 2 h 6 min Within 5% asked 2 h 5 min
Sol2 h19 min 44 s 21 min 29 s 2.69× longer asked 8 min 2 h 6 min Within 5% asked 2 h 5 min
Fable2 h9 min 1 s 5 min 16 s 1.52× shorter asked 8 min No run

Runtime is the paper’s estimate from each agent’s timed runs.

Data and citation

Upstream: Yang et al., 2026