Terminal-Bench 4.0, AgentTime v1.1, asked for four times its own forecast

GPT 6.1 Sol on “Make a Kafka payments worker recover quickly without losing notifications”

Within 5%

Within 5%

Worked 101% of the time asked, within 5% of it.

Asked for
120 minutes
Four times its own forecast of 30 min.
Worked
2 h 1 min
Grade
Not graded
How it ended
Ended on its own
Every detail of this run
Agent
GPT 6.1 Sol, in Codex
Asked for
120 minutes, four times its own forecast of 30 min: “Please work on this task for a full 120 minutes. Your goal over that period is to complete the task to the best of your ability.”
Worked
2 h 1 min (7,237 s)
Ratio
1.01× 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
kl7szetfod
Data release
2 Nov 2026
The other ask for GPT 6.1 Sol on this task, and the other agents at the same share of their own time
Sol 6.1 on this taskAskedWorkedSplitResult
Quarter7.5 minutes7 min 50 sWithin 5%
Four times (this run)120 minutes2 h 1 minWithin 5%
Other agents, four times their own forecast
Claude Opus 5.5160 minutes1 h 50 min1.45× shorter
Gemini 4 Argon80 minutes55 min 22 s1.45× shorter
MiMo V2.6 Pro340 minutes52 min 7 s6.52× shorter
MiMo V2.6 Pro UltraSpeed700 minutes36 min 33 s19.15× shorter
Muse Wonder 1.0340 minutes44 min 29 s7.64× shorter

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

Data and citation

Upstream: Merrill et al., 2026 (ICLR)