Terminal-Bench 4.0
Agents work on tasks inside a terminal. The score is the share of tasks resolved, with a 95% confidence interval.
- Unit
- % resolved
- Results
- 35 results23 Tokscale models matched, 23 in the source
- Last update
- Oct 6, 2026Newest publish date
- Retrieved
- Oct 21, 2026Source data through Oct 6, 2026
- License
- Terms not stated
Conditions. Each row is one agent and model pair at one reasoning effort, as submitted to the official leaderboard. Cost and tokens are for the full run.
Attribution. Terminal-Bench 4.0 official leaderboard, tbench.ai
Results
One row per model: its best reported result. The rank is the position in this list, by score; the source's own ranking can differ.
Tokscale usage counts accounts that used the model in public Tokscale submissions, Oct 14 – Nov 12, 2026. It is shown next to the score and does not explain it.
| Rank | Model | Score (% resolved) | Conditions | Cost and tokens | Date | Tokscale usage (30 days) | Source link |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Listed as Opus 5.5 | 64.8%Confidence interval 61.7–68.0% | Max effort · Claude Code330 trials | $4,719 per run8B tokens per run | Oct 6, 2026Published | 709active users | Source for Claude Opus 5.5 (opens in a new tab) |
| 2 | Claude Sonnet 5.5Listed as Sonnet 5.5 | 61.8%Confidence interval 58.9–64.8% | Max effort · Claude Code330 trials | $7,334 per run19.4B tokens per run | Oct 6, 2026Published | 555active users | Source for Claude Sonnet 5.5 (opens in a new tab) |
| 3 | 58.2%Confidence interval 55.1–61.3% | Max effort · Codex330 trials | $634 per run1.5B tokens per run | Oct 6, 2026Published | 660active users | Source for GPT-6.1 Sol (opens in a new tab) | |
| 3 | 58.2%Confidence interval 55.4–61.0% | Max effort · Codex330 trials | $3,267 per run1.5B tokens per run | Sep 3, 2026Published | 563active users | Source for GPT-6 Astra (opens in a new tab) | |
| 5 | Claude Fable 5.1Listed as Fable 5.1 | 57.9%Confidence interval 54.5–61.2% | XHigh effort · Claude Code330 trials | $4,872 per run2.3B tokens per run | Sep 17, 2026Published | 361active users | Source for Claude Fable 5.1 (opens in a new tab) |
| 6 | Claude Opus 5Listed as Opus 5 | 53.9%Confidence interval 50.8–57.1% | XHigh effort · Claude Code330 trials | $6,086 per run6.9B tokens per run | Sep 17, 2026Published | 163active users | Source for Claude Opus 5 (opens in a new tab) |
| 7 | 49.4%Confidence interval 46.2–52.6% | Max effort · Codex330 trials | $1,451 per run5B tokens per run | Oct 6, 2026Published | 246active users | Source for GPT-6 Sol (opens in a new tab) | |
| 8 | Claude Fable 5Listed as Fable 5 | 44.5%Confidence interval 40.7–48.4% | Max effort · Claude Code330 trials | $7,265 per run3.8B tokens per run | Aug 27, 2026Published | 41active users | Source for Claude Fable 5 (opens in a new tab) |
| 9 | 41.8%Confidence interval 38.6–45.1% | Max effort · Claude Code330 trials | $2,728 per run8.7B tokens per run | Aug 27, 2026Published | 139active users | Source for GLM-5.3 (opens in a new tab) | |
| 10 | 37.6%Confidence interval 34.0–41.1% | XHigh effort · Grok Build330 trials | $3,683 per run5.5B tokens per run | Sep 21, 2026Published | 131active users | Source for Grok 4.7 (opens in a new tab) | |
| 11 | 37.3%Confidence interval 33.5–41.0% | Max effort · Codex330 trials | $2,542 per run4.4B tokens per run | Aug 27, 2026Published | 217active users | Source for GPT-5.6 Sol (opens in a new tab) | |
| 12 | GLM-5.3 FlashListed as GLM-5.3-Flash | 35.8%Confidence interval 32.2–39.3% | No reasoning · Claude Code330 trials | $2,883 per run2.9B tokens per run | Oct 6, 2026Published | 170active users | Source for GLM-5.3 Flash (opens in a new tab) |
| 13 | Qwen3.8 Max 0902Listed as Qwen3.8-Max-0902 | 27.0%Confidence interval 23.2–30.8% | Max effort · Claude Code330 trials | $4,636 per run3B tokens per run | Oct 6, 2026Published | 1active users | Source for Qwen3.8 Max 0902 (opens in a new tab) |
| 14 | Claude Opus 4.8Listed as Opus 4.8 | 23.6%Confidence interval 20.1–27.2% | Max effort · Claude Code330 trials | $6,481 per run6.4B tokens per run | Aug 27, 2026Published | 222active users | Source for Claude Opus 4.8 (opens in a new tab) |
| 15 | 21.5%Confidence interval 18.3–24.8% | Max effort · Codex330 trials | $1,734 per run5.7B tokens per run | Aug 27, 2026Published | 120active users | Source for GPT-5.6 Terra (opens in a new tab) | |
| 16 | 20.3%Confidence interval 17.2–23.4% | High effort · Grok Build330 trials | $3,592 per run4B tokens per run | Aug 27, 2026Published | 41active users | Source for Grok 4.6 (opens in a new tab) | |
| 17 | 19.1%Confidence interval 15.7–22.4% | High effort · mini-SWE-agent330 trials | $1,829 per run17.2B tokens per run | Sep 3, 2026Published | 66active users | Source for Gemini 3.8 Flash (opens in a new tab) | |
| 18 | 17.3%Confidence interval 14.4–20.1% | Max effort · Codex330 trials | $347 per run11.6B tokens per run | Aug 27, 2026Published | 158active users | Source for GPT-5.6 Luna (opens in a new tab) | |
| 19 | 16.4%Confidence interval 13.6–19.1% | Max effort · Codex330 trials | $72.14 per run3.9B tokens per run | Oct 6, 2026Published | 433active users | Source for GPT-6 Luna (opens in a new tab) | |
| 20 | 14.5%Confidence interval 11.6–17.5% | XHigh effort · Muse Code330 trials | Cost not reported9B tokens | Oct 6, 2026Published | 21active users | Source for Muse Spark 1.3 (opens in a new tab) | |
| 21 | Claude Sonnet 5Listed as Sonnet 5 | 12.4%Confidence interval 9.4–15.5% | Max effort · Claude Code330 trials | $9,604 per run21.6B tokens per run | Aug 27, 2026Published | 138active users | Source for Claude Sonnet 5 (opens in a new tab) |
| 21 | 12.4%Confidence interval 9.8–15.0% | High effort · Grok Build330 trials | $2,094 per run3.4B tokens per run | Aug 27, 2026Published | 9active users | Source for Grok 4.5 (opens in a new tab) | |
| 23 | 11.2%Confidence interval 8.8–13.7% | High effort · mini-SWE-agent330 trials | $1,262 per run11.1B tokens per run | Sep 3, 2026Published | 4active users | Source for Gemini 3.7 Flash (opens in a new tab) |
Showing 23 of 23 models
Score by release date
Each dot is one model's best result on Terminal-Bench 4.0, placed at the model's release date. The dashed line is the best score among models released up to that date.
- Used on Tokscale in the last 30 days
- No Tokscale usage in the last 30 days
- Best score so far
Claude Opus 5.5: 64.8% · released Sep 22, 2026 · Max effort · Claude Code · 709 active users on Tokscale in the last 30 days
23 models, released May 28, 2026 to Sep 29, 2026. Highest: Claude Opus 5.5, 64.8%, released Sep 22, 2026. Use the left and right arrow keys to read each model in release order.0%20%40%60%80%
Show data
| Model | Released | Score | Tokscale users, 30 days |
|---|---|---|---|
| GPT-6.1 Sol | Sep 29, 2026 | 58.2%Max effort · Codex | 660 |
| Claude Sonnet 5.5 | Sep 28, 2026 | 61.8%Max effort · Claude Code | 555 |
| GPT-6 Luna | Sep 22, 2026 | 16.4%Max effort · Codex | 433 |
| GPT-6 Sol | Sep 22, 2026 | 49.4%Max effort · Codex | 246 |
| Claude Opus 5.5 | Sep 22, 2026 | 64.8%Max effort · Claude Code | 709 |
| Grok 4.7 | Sep 21, 2026 | 37.6%XHigh effort · Grok Build | 131 |
| GPT-6 Astra | Sep 3, 2026 | 58.2%Max effort · Codex | 563 |
| Muse Spark 1.3 | Sep 2, 2026 | 14.5%XHigh effort · Muse Code | 21 |
| Gemini 3.8 Flash | Sep 2, 2026 | 19.1%High effort · mini-SWE-agent | 66 |
| Qwen3.8 Max 0902 | Sep 2, 2026 | 27.0%Max effort · Claude Code | 1 |
| Claude Fable 5.1 | Sep 1, 2026 | 57.9%XHigh effort · Claude Code | 361 |
| GLM-5.3 Flash | Aug 26, 2026 | 35.8%No reasoning · Claude Code | 170 |
| GLM-5.3 | Aug 14, 2026 | 41.8%Max effort · Claude Code | 139 |
| Gemini 3.7 Flash | Aug 13, 2026 | 11.2%High effort · mini-SWE-agent | 4 |
| Grok 4.6 | Aug 12, 2026 | 20.3%High effort · Grok Build | 41 |
| Claude Opus 5 | Jul 24, 2026 | 53.9%XHigh effort · Claude Code | 163 |
| Grok 4.5 | Jul 16, 2026 | 12.4%High effort · Grok Build | 9 |
| Claude Sonnet 5 | Jun 30, 2026 | 12.4%Max effort · Claude Code | 138 |
| GPT-5.6 Luna | Jun 26, 2026 | 17.3%Max effort · Codex | 158 |
| GPT-5.6 Terra | Jun 26, 2026 | 21.5%Max effort · Codex | 120 |
| GPT-5.6 Sol | Jun 26, 2026 | 37.3%Max effort · Codex | 217 |
| Claude Fable 5 | Jun 9, 2026 | 44.5%Max effort · Claude Code | 41 |
| Claude Opus 4.8 | May 28, 2026 | 23.6%Max effort · Claude Code | 222 |
Method and matching
How results are run
- Each row is one agent and model pair at one reasoning effort, as submitted to the official leaderboard. Cost and tokens are for the full run.
- Only result rows are stored. No task content is copied.
- The site's data terms are not confirmed. Every score links back to the leaderboard.
- The same model can appear with several agents. The board shows the best reported pair and names the agent.
Matching to Tokscale models
- 35 of 35 results are matched to 23 Tokscale models: 35 by normalized name.
- No match is marked as reviewed by hand yet.
- Reasoning effort, thinking budget and agent are conditions of a result. They never change which model a result belongs to.
License and attribution
- Terms not stated, retrieved Oct 21, 2026.
- Terminal-Bench 4.0 official leaderboard, tbench.ai
Sources and attribution
Each score keeps its own source, unit and run conditions. Tokscale does not combine benchmarks into one score, and a missing score means the model was not evaluated, not that it scored zero.