Benchmark Almanac

Terminal-Bench 4.0

Claude Opus 5.5 64.8% leads since Sep 2026 In use

Agents complete 66 hard tasks in containerized terminals across software, ML, science, security, and operations. Test scripts verify each result, scored as resolution rate.

It removed 8 saturated or flawed tasks from Terminal-Bench 3.0 and set a flat 8-hour agent timeout.

Made by Stanford University , Harbor , and Laude Institute Released Aug 2026 Terminal tasks All tasks public Last in a launch post Oct 2026 Last model added Oct 2026 Scores refreshed 2 Nov 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
harbor-framework/terminal-bench 998 stars Dataset Leaderboard tbench.ai
23
Models scored on its own leaderboard
23
Launch posts list it, from 14 labs
11
System cards report it, from 6 labs

Claude Opus 5.5 leads with 64.8%

Best score per model from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

23 models. Scroll sideways for the rest.

  1. 64.8 Claude Opus 5.5
  2. 61.8 Claude Sonnet 5.5
  3. 58.2 GPT-6 Astra
  4. 58.2 GPT-6.1 Sol
  5. 57.9 Claude Fable 5.1
  6. 53.9 Claude Opus 5
  7. 49.4 GPT-6 Sol
  8. 44.5 Claude Fable 5
  9. 41.8 GLM-5.3
  10. 37.6 Grok 4.7
  11. 37.3 GPT-5.6 Sol
  12. 35.8 GLM-5.3-Flash
  13. 27.0 Qwen3.8 Max (0902)
  14. 23.6 Claude Opus 4.8
  15. 21.5 GPT-5.6 Terra
  16. 20.3 Grok 4.6
  17. 19.1 Gemini 3.8 Flash
  18. 17.3 GPT-5.6 Luna
  19. 16.4 GPT-6 Luna
  20. 14.6 Muse Spark 1.3
  21. 12.4 Claude Sonnet 5
  22. 12.4 Grok 4.5
  23. 11.2 Gemini 3.7 Flash
  1. 1 Claude Opus 5.5 max 64.8%
  2. 2 Claude Sonnet 5.5 max 61.8%
  3. 3 GPT-6 Astra max 58.2%
  4. 4 GPT-6.1 Sol max 58.2%
  5. 5 Claude Fable 5.1 max 57.9%
  6. 6 Claude Opus 5 xhigh 53.9%
  7. 7 GPT-6 Sol max 49.4%
  8. 8 Claude Fable 5 max 44.5%
  9. 9 GLM-5.3 max 41.8%
  10. 10 Grok 4.7 xhigh 37.6%
  11. 11 GPT-5.6 Sol max 37.3%
  12. 12 GLM-5.3-Flash none 35.8%
  13. 13 Qwen3.8 Max (0902) max 27%
  14. 14 Claude Opus 4.8 max 23.6%
  15. 15 GPT-5.6 Terra max 21.5%
  16. 16 Grok 4.6 high 20.3%
  17. 17 Gemini 3.8 Flash high 19.1%
  18. 18 GPT-5.6 Luna max 17.3%
  19. 19 GPT-6 Luna max 16.4%
  20. 20 Muse Spark 1.3 xhigh 14.6%
  21. 21 Claude Sonnet 5 max 12.4%
  22. 22 Grok 4.5 high 12.4%
  23. 23 Gemini 3.7 Flash high 11.2%
#ModelScoreSettingDateSource
1 Claude Opus 5.5 64.8% max, Claude Code Sep 2026 Official leaderboard
2 Claude Sonnet 5.5 61.8% max, Claude Code Sep 2026 Official leaderboard
3 GPT-6 Astra 58.2% max, Codex Sep 2026 Official leaderboard
4 GPT-6.1 Sol 58.2% max, Codex Sep 2026 Official leaderboard
5 Claude Fable 5.1 57.9% max, Claude Code Sep 2026 Official leaderboard
6 Claude Opus 5 53.9% xhigh, Claude Code Jul 2026 Official leaderboard
7 GPT-6 Sol 49.4% max, Codex Sep 2026 Official leaderboard
8 Claude Fable 5 44.5% max, Claude Code Jun 2026 Official leaderboard
9 GLM-5.3 41.8% max, Claude Code Aug 2026 Official leaderboard
10 Grok 4.7 37.6% xhigh, Grok Build Sep 2026 Official leaderboard
11 GPT-5.6 Sol 37.3% max, Codex Jul 2026 Official leaderboard
12 GLM-5.3-Flash 35.8% none, Claude Code Aug 2026 Official leaderboard
13 Qwen3.8 Max (0902) 27% max, Claude Code Sep 2026 Official leaderboard
14 Claude Opus 4.8 23.6% max, Claude Code May 2026 Official leaderboard
15 GPT-5.6 Terra 21.5% max, Codex Jul 2026 Official leaderboard
16 Grok 4.6 20.3% high, Grok Build Aug 2026 Official leaderboard
17 Gemini 3.8 Flash 19.1% high, mini-SWE-agent Sep 2026 Official leaderboard
18 GPT-5.6 Luna 17.3% max, Codex Jul 2026 Official leaderboard
19 GPT-6 Luna 16.4% max, Codex Sep 2026 Official leaderboard
20 Muse Spark 1.3 14.6% xhigh, Muse Code Sep 2026 Official leaderboard
21 Claude Sonnet 5 12.4% max, Claude Code Jun 2026 Official leaderboard
22 Grok 4.5 12.4% high, Grok Build Jul 2026 Official leaderboard
23 Gemini 3.7 Flash 11.2% high, mini-SWE-agent Aug 2026 Official leaderboard

Labs report scores for 23 models in their cards. We plot every source in the charts below.

More effort adds a median 15 points

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
GPT-6 Astra Codex
low medium high
58.2%
Claude Fable 5.1 Claude Code
low medium xhigh
57.9%
Claude Opus 5 Claude Code
low medium high xhigh max
53.9%
ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5.5 max Claude Code 64.8%
Claude Sonnet 5.5 max Claude Code 61.8%
GPT-6 Astra max Codex 58.2%
GPT-6.1 Sol max Codex 58.2%
Claude Fable 5.1 max Claude Code 57.9%
Claude Fable 5.1 xhigh Claude Code 57.9%
GPT-6 Astra xhigh Codex 57.9%
GPT-6 Astra high Codex 57.9%
Claude Fable 5.1 high Claude Code 54.5%
GPT-6 Astra medium Codex 54.2%
Claude Opus 5 xhigh Claude Code 53.9%
Claude Fable 5.1 medium Claude Code 53.9%
Claude Opus 5 max Claude Code 51.8%
GPT-6 Astra low Codex 50.6%
Claude Opus 5 high Claude Code 50.3%
GPT-6 Sol max Codex 49.4%
Claude Opus 5 medium Claude Code 44.9%
Claude Fable 5 max Claude Code 44.5%
Claude Fable 5.1 low Claude Code 43.3%
GLM-5.3 max Claude Code 41.8%
Grok 4.7 xhigh Grok Build 37.6%
GPT-5.6 Sol max Codex 37.3%
GLM-5.3-Flash none Claude Code 35.8%
Claude Opus 5 low Claude Code 34.9%
Qwen3.8 Max (0902) max Claude Code 27%
Claude Opus 4.8 max Claude Code 23.6%
GPT-5.6 Terra max Codex 21.5%
Grok 4.6 high Grok Build 20.3%
Gemini 3.8 Flash high mini-SWE-agent 19.1%
GPT-5.6 Luna max Codex 17.3%
GPT-6 Luna max Codex 16.4%
Muse Spark 1.3 xhigh Muse Code 14.6%
Claude Sonnet 5 max Claude Code 12.4%
Grok 4.5 high Grok Build 12.4%
Gemini 3.7 Flash high mini-SWE-agent 11.2%

4 models have held number one since its release in Aug 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

Anthropic OpenAI
Own leaderboard 4 leaders
Claude Opus 5.5 64.8%
Lab claims 3 leaders
Claude Sonnet 5.5 70.6%

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Sonnet 5.5 Claude Opus 5.5 Claude Fable 5.1 Claude Opus 5 Claude Fable 5 Claude Opus 4.8 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Sonnet 5.5 Claude Fable 5.1

14 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Anthropic
Google DeepMind
OpenAI
DeepSeek
Cognition
StepFun
xAI
Xiaomi Corp
Mistral AI
Meta
Ant Group
NVIDIA
MiniMax
Moonshot AI
DateLaunch postScoreAs the post writes it
29 Oct 2026 Mistral Large 4 (Le Chonk, open weights) 41 Terminal-Bench 4.0 (Artificial Analysis Coding Agent Index run)
28 Oct 2026 Gemini 4 Carbon 65.2 Terminal-Bench 4.0 (AA)
26 Oct 2026 GPT-6.1 Terra 50.3 Terminal-Bench 4.0, maximum reasoning effort
24 Oct 2026 Kimi K3.1 Pro 65.7 Terminal-Bench 4.0 (Claude Code)
19 Oct 2026 MiniMax-M3.1-Flash 38.1 Terminal Bench 4.0
17 Oct 2026 DeepSeek V4.1 Flash 1017 42.3 Terminal-Bench 4.0
15 Oct 2026 Nemotron 3.5 Super VL 1 Terminal Bench 4 [offline]
13 Oct 2026 Ling-3.1-flash 46 Terminal-Bench-4.0 (LingCode)
13 Oct 2026 Gemini 4 Argon 57.4 Terminal-bench 4.0
12 Oct 2026 Muse Wonder 52 Terminal-Bench 4.0-AA
12 Oct 2026 Muse Wonder Full 57.1 Terminal-Bench 4.0-AA
7 Oct 2026 Claude Haiku 5.5 39.2 Agentic coding Terminal-Bench 4.0
6 Oct 2026 Mistral Large 4 28.3 Coding - Terminal Bench 4.0: AA - Terminal-Bench 4 (chart 28; text 28.3%)
28 Sep 2026 Claude Sonnet 5.5 70.6 Agentic coding Terminal-Bench 4.0
22 Sep 2026 MiMo-V2.6 Pro 34.9 Terminal Bench 4.0
22 Sep 2026 Claude Opus 5.5 66.4 Agentic coding Terminal-Bench 4.0 (xhigh effort)
21 Sep 2026 Grok 4.7 37.6 Terminal-Bench 4.0 (Multi-hour terminal work)
20 Sep 2026 Step 5 Preview 33.3 Terminal-Bench v4
10 Sep 2026 SWE-2 27.3 Terminal-Bench 4.0
10 Sep 2026 DeepSeek V4.1 Flash 31.2 Terminal-Bench 4.0 (Pass@1)
3 Sep 2026 GPT-6 Astra 57.9 Terminal-Bench 4.0
2 Sep 2026 Gemini 3.8 Flash 19.1 Terminal-bench 4.0
1 Sep 2026 Claude Fable 5.1 60.9 Agentic coding Terminal-Bench 4.0 (Mythos 5.1)

11 system cards report it, from 6 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
13 Oct 2026 Gemini 4 Argon model card 57.4 Terminal-bench 4.0 (results table, p.5) Capabilities
12 Oct 2026 Muse Wonder model card 52 Terminal-Bench 4.0-AA accuracy from Artificial Analysis, Muse Wonder (max) (capabilities table image, page 3) Capabilities
12 Oct 2026 Muse Wonder Full model card 57.1 Terminal-Bench 4.0-AA accuracy from Artificial Analysis, Muse Wonder Full (max) (capabilities table image, page 3) Capabilities
7 Oct 2026 Claude Haiku 5.5 system card 39.2 Terminal-Bench 4.0, max effort, safeguards on with no fallback, no internet egress (8.4) Capabilities
28 Sep 2026 Claude Sonnet 5.5 system card 70.6 Terminal-Bench 4.0, max effort with safeguards and fallbacks, five trials per task (8.5) Capabilities
22 Sep 2026 MiMo-V2.6 Pro technical report 34.9 Terminal Bench 4.0 (Table 3) Capabilities
22 Sep 2026 Claude Opus 5.5 system card 66.4 Terminal-Bench 4.0 at xhigh effort, the setting the summary table marks (64.8% at max), Claude Code --bare mode, safeguards with fallback, five trials per task (Table 8.1.A, 8.5) Capabilities
21 Sep 2026 Grok 4.7 model card 38 Terminal-Bench 4.0, task success rate, xhigh, Grok Build harness, run by Harbor (2.3) Capabilities
10 Sep 2026 DeepSeek V4.1 Flash technical report 31.2 Terminal-Bench 4.0 (Pass@1) (Table 3) Capabilities
2 Sep 2026 Gemini 3.8 Flash model card 19.1 Terminal-bench 4.0 (results table, p.5) Capabilities
1 Sep 2026 Claude Fable 5.1 system card 55.8 Terminal-Bench 4.0, Claude Code --bare mode, max effort, 15 trials per task; Table 8.1.A rounds to 56% (Mythos 5.1 61%) (8.6) Capabilities

Versions

We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.

Anthropic OpenAI Google DeepMind Cognition
Terminal-Bench 4.0 20 leaders
Claude Sonnet 5.5 70.6%
Terminal-Bench Replaced by 2.0, Dec 2025
Claude Opus 4 50%
Terminal-Bench 2.0 Replaced by 2.1, Jul 2026
GPT-5.5 84.7%
Terminal-Bench 2.1 Passed 85% in Jun 2026, by Claude Fable 5
SWE-2 92.8%
Terminal-Bench 3.0 Replaced by 4.0, Sep 2026
Claude Opus 5 43.3%
Terminal-Bench 4.0 4 leaders
Claude Sonnet 5.5 70.6%
Replaces Terminal-Bench 3.0

More in Terminal agents

All 18 in Terminal agents
  • SkillsBench In use 86-87 containerized expert tasks in 11 domains, run with and without Agent Skills. Verifier tests give pass rate and skill lift. GPT-5.5 67.3%
  • Terminal-Bench Hard In use Agents complete 44 hard tasks from Terminal-Bench Core in the Terminus 2 harness. Artificial Analysis scores pass@1 averaged over 3 repeats. Gemma 4 31B IT 36%
  • Harbor-Index In use Agents solve about 80 hard tasks drawn from 29 benchmarks run through Harbor adapters, such as HLE, GSO and SWE-bench Pro. Each task keeps its source verifier. GPT-5.5 28.1%
  • CompileBench Not in model cards Agents build real open-source projects through dependency problems, legacy toolchains and complex build systems, for example compiling 22-year-old code. Claude Opus 4.1 100%
  • Long-Horizon-Terminal-Bench 2.0 Not in model cards Agents work for up to 3 hours per task on 77 container tasks, including 30 tasks that rebuild a scientific tool from its outputs. Hidden verifiers give partial-credit rewards. GPT-6 Astra 69.2%
  • ITBench Not in model cards Agents handle IT automation tasks in Kubernetes environments: site reliability incidents, compliance and security checks, and cloud cost work. Gemini 3 Pro 58.3%