BenchmarksTerminal-Bench 4.0 cost per task

Terminal-Bench 4.0 cost per task

Average USD cost of an agent attempt on Terminal-Bench 4.0 at a stated reasoning effort.

Inference costRecipe not catalogued

Reproduce this test ↗Author page ↗

Test definition

Author / evaluator: Reported by Moonshot AI
Version / family: Terminal-Bench 4.0 cost per task / Terminal-Bench 4.0 cost per task
Evidence available: 16 results in 5 comparison groups

How the test is scored

The official score/cost chart quotes tbench.ai measurements except DeepSeek, which quotes Artificial Analysis. Each effort remains its own row.

Comparison notes

Lower USD cost is less spending per attempt, not higher quality or successful-task efficiency. Read with the paired task-success score.

A different task split, tool policy, judge or reasoning effort can change the result. Choose one comparison group below.

Results by evaluator

Select a publication and metric to compare the recorded field.

Comparison group

selfTerminal-Bench 4.0: complete score/cost chart · Terminal-Bench 4.0 cost per task (low)

selfUSD · Lower is better

1GPT-6 Astra4.72
2Claude Fable 5.17.15

Bars start at zero. 2 reported results in this comparison group; ties share a rank. Source ↗

All observations

Every numeric result, with its source and reported run conditions.

Find a resultSource

16 results

Results (USD)
Model	Result (USD)	Source	Run
Claude Fable 5.1Lower is better	7.15low	self	Details Average USD per task at low effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (low) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	8.58medium	self	Details Average USD per task at medium effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (medium) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	12.08high	self	Details Average USD per task at high effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (high) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	14.76xhigh	self	Details Average USD per task at xhigh effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (xhigh) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	18.92max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
GPT-6 AstraLower is better	4.72low	self	Details Average USD per task at low effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (low) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	5.8medium	self	Details Average USD per task at medium effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (medium) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	6.88high	self	Details Average USD per task at high effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (high) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	7.12xhigh	self	Details Average USD per task at xhigh effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (xhigh) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
	9.9max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
Kimi K3.1 ProLower is better	4.16high	self	Details Average USD per task at high effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (high) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
Kimi K3.1 ProLower is better	5.25max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
Claude Sonnet 5.5Lower is better	22.22max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
Claude Opus 5.5Lower is better	14.3max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
GLM 5.3Lower is better	8.27max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →
DeepSeek V4.1 Flash 1017Lower is better	0.88max	self	Details Average USD per task at max effort, quoted by Moonshot from tbench.ai (DeepSeek from AA). Lower cost does not imply higher task success; use the paired score. Exact numeric coordinates from the official embedded chart, read 24 October 2026. All 16 points retain both score and USD cost plus their effort. Data quoted from tbench.ai except DeepSeek, which quotes AA. Publisher-chart values stay separate from directly imported independent boards and the headline release comparison. Metric Terminal-Bench 4.0 cost per task (max) Published 23 October 2026 Effort Not reported separately Tasks / attempts Unknown / Unknown Confidence interval Not reported Build a reproduction plan →

Reproduction

Recipe not catalogued

Public execution instructions have not been verified in this catalogue. This status does not claim that the benchmark is closed or impossible to run.

Execution documentation ↗

Build a run plan →