Agents and computer useOfficial + Vendor scoresPrices verified
Terminal-Bench 4.0 leaderboard
As of November 8, 2026, Kimi K3.1 Pro (Moonshot) leads Terminal-Bench 4.0 at 65.7%, ahead of Claude Opus 5.5 at 64.8% and Claude Sonnet 5.5 at 61.8%. 33 systems ranked. 23 of the 33 scores on the main board come from Terminal-Bench (Stanford, Laude Institute); the other 10 are labs' reports on their own models, linked on every row.
Top score
65.7%
Systems ranked
33
Independent rows
23/45
Cheapest near top
$1.28/1M
With API prices
33/33
Newest score
Nov 2
What is Terminal-Bench 4.0? Terminal-Bench 4.0 is the current Terminal-Bench release: the 3.0 task format with fixed tasks, saturated tasks removed and resources recalibrated, run on an official leaderboard. Made by The Terminal-Bench and Harbor team (Laude Institute), August 2026.
§ 01 / LEADERBOARD
Top models on Terminal-Bench 4.0, November 2026
| # | Model | Score | $ / task | Price / 1M | Context | Source |
|---|---|---|---|---|---|---|
| 1 | Kimi K3.1 Pro Moonshot (Kimi) | max effort | — | $2 / $10 | 1M | Vendor-reported Moonshot AI |
| 2 | Claude Opus 5.5 Anthropic | max effort, Claude Code harness, 330 trials, ±3.1 pts (95% CI) | $14.30 | $4 / $20 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 3 | Claude Sonnet 5.5 Anthropic | max effort, Claude Code harness, 330 trials, ±2.9 pts (95% CI) | $22.22 | $2 / $10 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 4 | GPT-6 Astra OpenAI | max effort, Codex harness, 330 trials, ±2.8 pts (95% CI) | $9.90 | $10 / $50 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 4 | GPT-6.1 Sol OpenAI | max effort, Codex harness, 330 trials, ±3.1 pts (95% CI) | $1.92 | $2 / $10 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 6 | Claude Fable 5.1 Anthropic | xhigh effort, Claude Code harness, 330 trials, ±3.4 pts (95% CI) | $14.76 | $10 / $50 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 7 | Gemini 4 Argon Google DeepMind | as reported | — | $2 / $10 | 1M | Vendor-reported Google DeepMind |
| 8 | Claude Opus 5 Anthropic | xhigh effort, Claude Code harness, 330 trials, ±3.2 pts (95% CI) | $18.44 | $5 / $25 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 9 | GPT-6 Sol OpenAI | max effort, Codex harness, 330 trials, ±3.2 pts (95% CI) | $4.40 | $2 / $10 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 10 | Claude Fable 5 Anthropic | max effort, Claude Code harness, 330 trials, ±3.9 pts (95% CI) | $22.02 | $10 / $50 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 11 | DeepSeek V4.1 Flash DeepSeek · listed as “DeepSeek V4.1 Flash (1017)” | as reported | — | $0.3 / $1.2 | 1M | Vendor-reported DeepSeek |
| 12 | GLM-5.3 Zhipu (Z.ai / GLM) | max effort, Claude Code harness, 330 trials, ±3.2 pts (95% CI) | $8.27 | $1.4 / $4.4 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 13 | Claude Haiku 5.5 Anthropic | best reported effort | — | $0.1 / $0.5 | 1M | Vendor-reported Anthropic |
| 14 | MiniMax M3.1 Flash MiniMax | as reported | — | $0.15 / $0.6 | 1M | Vendor-reported MiniMax |
| 15 | Grok 4.7 SpaceXSI | xhigh effort, Grok Build harness, 330 trials, ±3.5 pts (95% CI) | $11.16 | $2 / $6 | 500K | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 16 | GPT-5.6 Sol OpenAI | max effort, Codex harness, 330 trials, ±3.8 pts (95% CI) | $7.70 | $4 / $20 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 17 | GLM-5.3-Flash Zhipu (Z.ai / GLM) | no reasoning effort set, Claude Code harness, 330 trials, ±3.5 pts (95% CI) | $0.36 | $0.15 / $0.5 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 18 | MiMo-V2.6-Pro Xiaomi MiMo | as reported | — | $0.435 / $0.87 | 1M | Vendor-reported Xiaomi |
| 19 | MiMo-V2.6-Flash Xiaomi MiMo | as reported | — | $0.14 / $0.28 | 1M | Vendor-reported Xiaomi |
| 20 | Qwen3.8 Max Alibaba (Qwen) · listed as “Qwen3.8 Max (0902)” | max effort, Claude Code harness, 330 trials, ±3.8 pts (95% CI) | $14.05 | $2 / $6 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 21 | Claude Opus 4.8 Anthropic | max effort, Claude Code harness, 330 trials, ±3.6 pts (95% CI) | $19.64 | $5 / $25 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 22 | GPT-5.6 Terra OpenAI | max effort, Codex harness, 330 trials, ±3.3 pts (95% CI) | $5.25 | $2 / $12 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 23 | Grok 4.6 SpaceXSI | high effort, Grok Build harness, 330 trials, ±3.1 pts (95% CI) | $10.88 | $2 / $6 | 500K | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 24 | Gemini 3.8 Flash Google DeepMind | high effort, mini-SWE-agent harness, 330 trials, ±3.4 pts (95% CI) | $5.54 | $0.75 / $3.75 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 25 | GPT-5.6 Luna OpenAI | max effort, Codex harness, 330 trials, ±2.9 pts (95% CI) | $1.05 | $0.2 / $1.2 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 26 | GPT-6 Luna OpenAI | max effort, Codex harness, 330 trials, ±2.7 pts (95% CI) | $0.22 | $0.1 / $0.5 | 1.05M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 27 | Muse Spark 1.3 Meta | xhigh effort, Muse Code harness, 330 trials, ±2.9 pts (95% CI) | — | $1.25 / $4.25 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 28 | Claude Sonnet 5 Anthropic | max effort, Claude Code harness, 330 trials, ±3.1 pts (95% CI) | $29.10 | $2 / $10 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 28 | Grok 4.5 SpaceXSI | high effort, Grok Build harness, 330 trials, ±2.6 pts (95% CI) | $6.35 | $2 / $6 | 500K | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 28 | DeepSeek V4 Pro DeepSeek | pass@1 | — | $1.32 / $3.96 | 1M | Vendor-reported DeepSeek |
| 31 | Gemini 3.7 Flash Deprecated Google DeepMind | high effort, mini-SWE-agent harness, 330 trials, ±2.5 pts (95% CI) | $3.82 | $0.75 / $3.75 | 1M | Official leaderboard Terminal-Bench (Stanford, Laude Institute) |
| 32 | DeepSeek V4 Flash Deprecated DeepSeek | pass@1 | — | $0.44 / $1.32 | 1M | Vendor-reported DeepSeek |
| 33 | Claude Haiku 4.5 Anthropic | as reported | — | $1 / $5 | 200K | Vendor-reported Anthropic |
§ 02 / SCORE VS PRICE
Terminal-Bench 4.0: score against price
§ 03 / REASONING EFFORT
What more reasoning buys
Some publishers ran the same model at several reasoning-effort levels and printed the score and the cost per task for each. Higher effort spends more tokens per task; this is what it bought on Terminal-Bench 4.0.
| Model | Score by effort level ($ per task) | Source |
|---|---|---|
| GPT-6 Astra | low 50.6% $4.72 medium 54.2% $5.80 high 57.9% $6.88 xhigh 57.9% $7.12 max 58.2% $9.90 | Terminal-Bench (Stanford, Laude Institute) |
| Claude Fable 5.1 | low 43.3% $7.15 medium 53.9% $8.58 high 54.5% $12.08 xhigh 57.9% $14.76 max 57.9% $18.92 | Terminal-Bench (Stanford, Laude Institute) |
| Claude Opus 5 | low 34.9% $7.25 medium 44.9% $9.67 high 50.3% $14.13 xhigh 53.9% $18.44 max 51.8% $18.09 | Terminal-Bench (Stanford, Laude Institute) |
§ 04 / PROVENANCE
Can you trust Terminal-Bench 4.0 scores?
23 scores on this page come from the benchmark's own leaderboard (Terminal-Bench (Stanford, Laude Institute)), where every model runs on the same harness. The other 22 are labs' reports on their own models. 11 models have numbers from two sources for the same setting. For 5 of them the numbers differ by 1 point or more; they are listed below, and the table ranks the independent number.
Where the sources disagree
- Claude Sonnet 5.5 8.8 pts apart 61.8% Terminal-Bench (Stanford, Laude Institute), as reported · 2026-09-28 70.6% Anthropic, as reported · 2026-09-28
- Claude Fable 5.1 2.1 pts apart 57.9% Terminal-Bench (Stanford, Laude Institute), as reported · 2026-09-01 55.8% Anthropic, as reported · 2026-09-22
- Claude Sonnet 5 2.1 pts apart 12.4% Terminal-Bench (Stanford, Laude Institute), as reported · 2026-06-30 10.3% Anthropic, as reported · 2026-09-28
- Claude Opus 5 1.6 pts apart 53.9% Terminal-Bench (Stanford, Laude Institute), as reported · 2026-07-24 52.3% Anthropic, as reported · 2026-09-22
- Claude Opus 5.5 1.6 pts apart 64.8% Terminal-Bench (Stanford, Laude Institute), as reported · 2026-09-22 66.4% Anthropic, as reported · 2026-09-22
- 4.0 is a breaking change from 3.0: tasks were removed and fixed, so scores from the two versions are not comparable.
- Every task has a flat 8-hour timeout; the remaining errors are mostly refusals and output-token limits.
- Saturated tasks were pruned using the latest generation of models, so the set is tuned to today's frontier.
§ 05 / HOW IT WORKS
How Terminal-Bench 4.0 works
The task
An agent in a container solves a hard terminal job in engineering, science, data or operations, and the files it produces are tested.
The score
Resolution rate: the share of tasks whose verifier passes. The official leaderboard runs each model many times and publishes the rate with a 95% confidence interval, the cost of the run and the reasoning effort used.
Test set
The authors do not publish a single task count for this version.
Made by
The Terminal-Bench and Harbor team (Laude Institute), first published August 2026.
Official sources
- www.tbench.ai
- Definition read on 2026-11-08
§ 06 / SOURCES
Sources and freshness
45 published scores from 13 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.
- Official leaderboard Terminal-Bench (Stanford, Laude Institute)
- Vendor-reported Anthropic
- Vendor-reported DeepSeek
- Vendor-reported Anthropic
- Vendor-reported Anthropic
- Vendor-reported Google DeepMind
- Vendor-reported OpenAI
- Vendor-reported SpaceXSI (xAI)
- Vendor-reported Xiaomi
- Vendor-reported DeepSeek
- Vendor-reported Google DeepMind
- Vendor-reported MiniMax
- Vendor-reported Moonshot AI
Terminal-Bench 4.0 FAQ
Answers derived from the 45 published scores on this page and the current price index.
Q · 01 Which AI model has the highest Terminal-Bench 4.0 score?
+
Kimi K3.1 Pro with 65.7% (max effort), as reported by Moonshot AI on November 2, 2026. Next: Claude Opus 5.5 at 64.8% and Claude Sonnet 5.5 at 61.8%.
Q · 02 What is the cheapest model near the top of Terminal-Bench 4.0?
+
The leader itself. No priced model within 5 points of Kimi K3.1 Pro is cheaper: it runs at $1.28 per 1M tokens on our effective blend ($2 input, $10 output).
Q · 03 Are these Terminal-Bench 4.0 scores independently verified?
+
23 of 45 published scores come from Terminal-Bench (Stanford, Laude Institute); the other 22 are vendor reports. Each row carries a badge that says which, and a link to the source.
Q · 04 What does Terminal-Bench 4.0 measure?
+
Terminal-Bench 4.0 is the current Terminal-Bench release: the 3.0 task format with fixed tasks, saturated tasks removed and resources recalibrated, run on an official leaderboard. An agent in a container solves a hard terminal job in engineering, science, data or operations, and the files it produces are tested.
Q · 05 How is the Terminal-Bench 4.0 score calculated?
+
Resolution rate: the share of tasks whose verifier passes. The official leaderboard runs each model many times and publishes the rate with a 95% confidence interval, the cost of the run and the reasoning effort used.
Q · 06 How current is this Terminal-Bench 4.0 leaderboard?
+
The newest score was published on November 2, 2026. Every row shows the date its source published it, and we last read all of them on November 8, 2026. Prices come from our model price index, where each rate carries the date it was read on the vendor's page.