Benchmark Almanac

Terminal-Bench 2.0

GPT-5.5 84.7% leads since Apr 2026 In use

89 hard terminal tasks across 16 categories in containers. Hidden tests verify the final state. 5 trials per task.

Leaderboard entries are audited and trajectories are public. Later versions exist (Vals now uses Terminal-Bench 4.0).

In the authors’ words

AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 2.0: a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 65% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work at https://www.tbench.ai/ .

Abstract of Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces (arXiv:2601.11868, CC0)
Made by Stanford + Laude Institute Released Nov 2025 Terminal tasks All tasks public Last in a launch post Jul 2026 Last model added Jul 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-2 438 stars Paper Dataset Leaderboard
46
Entries on its own leaderboard, 43 models we track
59
Launch posts list it, from 21 labs
28
System cards report it, from 17 labs

GPT-5.5 leads with 84.7% among models released by Apr 2026

Best score per entry from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

The newest model with a score here came out in Apr 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

46 entries. Scroll sideways for the rest.

  1. 84.7 GPT-5.5
  2. 80.2 Gemini 3.1 Pro
  3. 80.2 Claude Opus 4.7
  4. 78.4 GPT-5.3 Codex
  5. 76.4 Claude Opus 4.6
  6. 69.4 Gemini 3 Pro
  7. 66.5 GPT-5.2 Codex
  8. 64.9 GPT-5.2
  9. 64.3 Gemini 3 Flash
  10. 63.1 Claude Opus 4.5
  11. 61.6 GPT-5.1-Codex-mini
  12. 60.5 GPT-5.1-Codex-Max
  13. 57.8 GPT-5.1-Codex
  14. 57.3 Grok 4.20
  15. 53.4 Claude Sonnet 4.6
  16. 52.4 GLM-5
  17. 49.6 GPT-5
  18. 47.6 GPT-5.1
  19. 46.5 Claude Sonnet 4.5
  20. 45.1 MiniMax-M2.7
  21. 44.3 GPT-5-Codex
  22. 43.2 Kimi K2.5
  23. 42.9 MiniMax M2.7 Highspeed
  24. 42.7 MiniMax-M2.5
  25. 39.5 DeepSeek-V3.2
  26. 38.0 Claude Opus 4.1
  27. 36.6 MiniMax-M2.1
  28. 35.7 Kimi K2 Thinking
  29. 35.5 Claude Haiku 4.5
  30. 34.8 GPT-5 mini
  31. 33.4 GLM-4.7
  32. 32.6 Gemini 2.5 Pro
  33. 30.0 MiniMax-M2
  34. 27.8 Kimi K2
  35. 27.2 Grok 4
  36. 27.2 Qwen3-Coder-480B-A35B
  37. 25.8 grok-code-fast-1
  38. 24.6 Qwen 3.6 35B-A3B
  39. 24.5 GLM-4.6
  40. 21.8 GPT-5 nano
  41. 19.3 TermiGen-32B
  42. 18.7 gpt-oss-120b
  43. 17.1 Gemini 2.5 Flash
  44. 17.0 AfterQuery-GPT-OSS-20B
  45. 9.21 Qwen3.5-9B
  46. 3.37 gpt-oss-20b
  1. 1 GPT-5.5 84.7%
  2. 2 Gemini 3.1 Pro 80.2%
  3. 3 Claude Opus 4.7 80.2%
  4. 4 GPT-5.3 Codex 78.4%
  5. 5 Claude Opus 4.6 76.4%
  6. 6 Gemini 3 Pro 69.4%
  7. 7 GPT-5.2 Codex 66.5%
  8. 8 GPT-5.2 64.9%
  9. 9 Gemini 3 Flash 64.3%
  10. 10 Claude Opus 4.5 63.1%
  11. 11 GPT-5.1-Codex-mini 61.6%
  12. 12 GPT-5.1-Codex-Max 60.5%
  13. 13 GPT-5.1-Codex 57.8%
  14. 14 Grok 4.20 reasoning 57.3%
  15. 15 Claude Sonnet 4.6 53.4%
  16. 16 GLM-5 52.4%
  17. 17 GPT-5 49.6%
  18. 18 GPT-5.1 47.6%
  19. 19 Claude Sonnet 4.5 46.5%
  20. 20 MiniMax-M2.7 45.1%
  21. 21 GPT-5-Codex 44.3%
  22. 22 Kimi K2.5 43.2%
  23. 23 MiniMax M2.7 Highspeed 42.9%
  24. 24 MiniMax-M2.5 42.7%
  25. 25 DeepSeek-V3.2 39.5%
  26. 26 Claude Opus 4.1 38%
  27. 27 MiniMax-M2.1 36.6%
  28. 28 Kimi K2 Thinking 35.7%
  29. 29 Claude Haiku 4.5 35.5%
  30. 30 GPT-5 mini 34.8%
  31. 31 GLM-4.7 33.4%
  32. 32 Gemini 2.5 Pro Jun 2025 32.6%
  33. 33 MiniMax-M2 30%
  34. 34 Kimi K2 Sep 2025 27.8%
  35. 35 Grok 4 27.2%
  36. 36 Qwen3-Coder-480B-A35B 27.2%
  37. 37 grok-code-fast-1 25.8%
  38. 38 Qwen 3.6 35B-A3B 24.6%
  39. 39 GLM-4.6 24.5%
  40. 40 GPT-5 nano 21.8%
  41. 41 TermiGen-32B 19.3%
  42. 42 gpt-oss-120b 18.7%
  43. 43 Gemini 2.5 Flash Jun 2025 17.1%
  44. 44 AfterQuery-GPT-OSS-20B 17%
  45. 45 Qwen3.5-9B 9.21%
  46. 46 gpt-oss-20b 3.37%
#EntryScoreSettingDateSource
1 GPT-5.5 84.7% NexAU-AHE Apr 2026 Official leaderboard
2 Gemini 3.1 Pro 80.2% TongAgents Feb 2026 Official leaderboard
3 Claude Opus 4.7 80.2% WOZCODE Apr 2026 Official leaderboard
4 GPT-5.3 Codex 78.4% SageAgent Feb 2026 Official leaderboard
5 Claude Opus 4.6 76.4% Meta-Harness Feb 2026 Official leaderboard
6 Gemini 3 Pro 69.4% Ante Nov 2025 Official leaderboard
7 GPT-5.2 Codex 66.5% Deep Agents Dec 2025 Official leaderboard
8 GPT-5.2 64.9% Droid Dec 2025 Official leaderboard
9 Gemini 3 Flash 64.3% Junie CLI Dec 2025 Official leaderboard
10 Claude Opus 4.5 63.1% Droid Nov 2025 Official leaderboard
11 GPT-5.1-Codex-mini 61.6% hookele Nov 2025 Official leaderboard
12 GPT-5.1-Codex-Max 60.5% Codex CLI Nov 2025 Official leaderboard
13 GPT-5.1-Codex 57.8% Crux Nov 2025 Official leaderboard
14 Grok 4.20 57.3% reasoning, Grok CLI Feb 2026 Official leaderboard
15 Claude Sonnet 4.6 53.4% Simplai Agent Feb 2026 Official leaderboard
16 GLM-5 52.4% Terminus 2 Feb 2026 Official leaderboard
17 GPT-5 49.6% Codex CLI Aug 2025 Official leaderboard
18 GPT-5.1 47.6% Terminus 2 Nov 2025 Official leaderboard
19 Claude Sonnet 4.5 46.5% CAMEL-AI Sep 2025 Official leaderboard
20 MiniMax-M2.7 45.1% IndusAGI Coding Agent Mar 2026 Official leaderboard
21 GPT-5-Codex 44.3% Codex CLI Sep 2025 Official leaderboard
22 Kimi K2.5 43.2% Terminus 2 Jan 2026 Official leaderboard
23 MiniMax M2.7 Highspeed 42.9% Harness Agent Mar 2026 Official leaderboard
24 MiniMax-M2.5 42.7% cchuter Feb 2026 Official leaderboard
25 DeepSeek-V3.2 39.5% Terminus 2 Dec 2025 Official leaderboard
26 Claude Opus 4.1 38% Terminus 2 Aug 2025 Official leaderboard
27 MiniMax-M2.1 36.6% Crux Dec 2025 Official leaderboard
28 Kimi K2 Thinking 35.7% Terminus 2 Nov 2025 Official leaderboard
29 Claude Haiku 4.5 35.5% Goose Oct 2025 Official leaderboard
30 GPT-5 mini 34.8% spoox-o-m Aug 2025 Official leaderboard
31 GLM-4.7 33.4% Terminus 2 Dec 2025 Official leaderboard
32 Gemini 2.5 Pro Jun 2025 32.6% Terminus 2 Jun 2025 Official leaderboard
33 MiniMax-M2 30% Terminus 2 Oct 2025 Official leaderboard
34 Kimi K2 Sep 2025 27.8% Terminus 2 Sep 2025 Official leaderboard
35 Grok 4 27.2% OpenHands Jul 2025 Official leaderboard
36 Qwen3-Coder-480B-A35B 27.2% Dakou Agent Jul 2025 Official leaderboard
37 grok-code-fast-1 25.8% Mini-SWE-Agent Aug 2025 Official leaderboard
38 Qwen 3.6 35B-A3B 24.6% little-coder Apr 2026 Official leaderboard
39 GLM-4.6 24.5% Terminus 2 Sep 2025 Official leaderboard
40 GPT-5 nano 21.8% spoox-o-m Aug 2025 Official leaderboard
41 TermiGen-32B 19.3% Bash Agent May 2026 Official leaderboard
42 gpt-oss-120b 18.7% Terminus 2 Aug 2025 Official leaderboard
43 Gemini 2.5 Flash Jun 2025 17.1% Mini-SWE-Agent Jun 2025 Official leaderboard
44 AfterQuery-GPT-OSS-20B 17% Terminus 2 Mar 2026 Official leaderboard
45 Qwen3.5-9B 9.21% little-coder Feb 2026 Official leaderboard
46 gpt-oss-20b 3.37% Mini-SWE-Agent Aug 2025 Official leaderboard

Epoch AI has scores for 43 models. Labs report scores for 61 models in their cards. We plot every source in the charts below.

6 models have held number one since its release in Nov 2025

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, Epoch AI’s runs and what labs claim in their cards.

OpenAI Google DeepMind Anthropic
Own leaderboard Replaced by 2.1, May 2026
GPT-5.5 84.7%
Epoch AI runs Replaced by 2.1, May 2026
GPT-5.5 84.7%
Lab claims Replaced by 2.1, Jul 2026
GPT-5.5 82.7%

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 Run or collected by Epoch AI Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-5.5 Gemini 3.1 Pro Gemini 3 Pro GPT-5.1-Codex-mini GPT-5 Claude Opus 4.1 Gemini 2.5 Pro 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-5.5 Gemini 3 Pro

4 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.

Google DeepMind
OpenAI
Anthropic
DeepSeek
Mistral AI
Xiaomi Corp
ByteDance Seed
Z.ai
MiniMax
LG AI Research
Moonshot AI
StepFun
Qwen
Cursor
NVIDIA
Upstage
Meta
Tencent
Poolside
Microsoft
Cohere
DateLaunch postScoreAs the post writes it
2 Jul 2026 Laguna XS 2.1 37.5 Terminal-Bench 2.0
9 Jun 2026 North Mini Code 1.0 36 Terminal-Bench v2 (simple ReAct harness with one terminal tool)
2 Jun 2026 MAI-Code-1-Flash 54.8 Terminal Bench 2
2 Jun 2026 MAI-Thinking-1 46 Terminal-Bench 2.0
1 Jun 2026 Qwen3.7-Plus 70.3 Terminal Bench 2.0-Terminus (avg of 5 runs)
20 May 2026 Qwen3.7-Max 69.7 Terminal Bench 2.0-Terminus
18 May 2026 Composer 2.5 69.3 Terminal-Bench 2.0
28 Apr 2026 Laguna XS.2 30.1 Terminal-Bench 2.0 (mean pass@1 over 5 runs)
28 Apr 2026 Laguna M.1 40.7 Terminal-Bench 2.0 (mean pass@1 over 5 runs, 48GB RAM/32 CPUs)
28 Apr 2026 Seed2.0 Lite (0428) 43.3 Terminal Bench 2.0
27 Apr 2026 MiMo-V2.5-Pro 68.4 Terminal-Bench 2.0
24 Apr 2026 DeepSeek-V4-Pro 67.9 Terminal Bench 2.0 (Acc)
23 Apr 2026 Tencent Hy3 preview 54.4 Terminal-Bench 2.0
23 Apr 2026 GPT-5.5 82.7 Terminal-Bench 2.0
22 Apr 2026 Qwen3.6 27B 59.3 Terminal-Bench 2.0 (Terminus-2, avg of 5)
22 Apr 2026 MiMo-V2.5 65.8 Terminal-Bench 2.0
20 Apr 2026 Qwen 3.6 Max (Preview) 65.4 Terminal-Bench 2.0 (Terminus-2) Agentic Terminal Coding
20 Apr 2026 Kimi K2.6 66.7 Terminal-Bench 2.0 (Terminus-2)
16 Apr 2026 Claude Opus 4.7 69.4 Agentic terminal coding Terminal-Bench 2.0
15 Apr 2026 Qwen 3.6 35B-A3B 51.5 Terminal-Bench 2.0 (Terminus-2, avg of 5)
8 Apr 2026 Muse Spark 59 Terminal-Bench 2.0 (Agentic Terminal Coding)
7 Apr 2026 GLM-5.1 69 Terminal-Bench 2.0 (Best self-reported), Claude Code
7 Apr 2026 Claude Mythos Preview 82 Terminal-Bench 2.0 (Terminus-2)
2 Apr 2026 Qwen 3.6 Plus 61.6 Terminal-Bench 2.0
24 Mar 2026 Solar Pro 3 (March 2026 update) 10.1 Terminal Bench 2
19 Mar 2026 Nemotron-Cascade 2 30B-A3B 21.1 Terminal Bench 2.0
19 Mar 2026 Composer 2 61.7 Terminal-Bench 2.0 (Harbor, mean of 5 runs)
18 Mar 2026 MiMo-V2-Pro 57.1 Terminal-Bench 2.0
18 Mar 2026 MiniMax-M2.7 57 Terminal Bench 2 (text)
17 Mar 2026 GPT-5.4 Nano 46.3 Terminal-Bench 2.0
17 Mar 2026 GPT-5.4 Mini 60 Terminal-Bench 2.0
11 Mar 2026 Nemotron 3 Super 31 Terminal Bench Core 2.0
24 Feb 2026 Qwen3.5-122B-A10B 49.4 Terminal Bench 2
19 Feb 2026 Gemini 3.1 Pro 68.5 Terminal-Bench 2.0 (Terminus-2 harness)
17 Feb 2026 Claude Sonnet 4.6 59.1 Agentic terminal coding Terminal-Bench 2.0
16 Feb 2026 Qwen3.5 397B-A17B 52.5 Terminal Bench 2
14 Feb 2026 Seed2.0 Pro 55.8 Terminal Bench 2.0 (Terminus 2, 3 tasks excluded)
12 Feb 2026 GPT-5.3-Codex-Spark 58.4 Terminal-Bench 2.0
12 Feb 2026 MiniMax-M2.5 51.7 Terminal Bench 2
11 Feb 2026 GLM-5 56.2 Terminal-Bench 2.0 (Terminus 2)
9 Feb 2026 Composer 1.5 47.9 Terminal-Bench 2.0 Score (%) (Harbor, mean of 2 runs)
5 Feb 2026 Claude Opus 4.6 65.4 Agentic terminal coding Terminal-Bench 2.0
5 Feb 2026 GPT-5.3 Codex 77.3 Terminal-Bench 2.0
2 Feb 2026 qwen3-coder-next 36.2 Terminal-Bench 2.0 (w/ Terminus-2 json)
2 Feb 2026 Step 3.5 Flash 51 Terminal-Bench 2.0
27 Jan 2026 Kimi K2.5 50.8 Terminal Bench 2.0
31 Dec 2025 K-EXAONE 29 Terminal-Bench 2.0
23 Dec 2025 MiniMax-M2.1 47.9 Terminal-bench 2.0
22 Dec 2025 GLM-4.7 41 Terminal Bench 2.0
18 Dec 2025 GPT-5.2 Codex 64 Terminal-Bench 2.0
18 Dec 2025 Seed1.8 45.2 Terminal Bench 2.0
17 Dec 2025 Gemini 3 Flash 47.6 Terminal-bench 2.0 (Terminus-2 harness)
16 Dec 2025 MiMo-V2-Flash 38.5 Terminal-Bench 2.0
9 Dec 2025 Devstral Small 2 22.5 Terminal Bench 2 (HF card)
9 Dec 2025 Devstral 2 32.6 Terminal Bench 2 (HF card)
1 Dec 2025 DeepSeek-V3.2 46.4 Terminal Bench 2.0 (Acc)
24 Nov 2025 Claude Opus 4.5 59.3 Agentic terminal coding Terminal-bench 2.0 (128k thinking budget)
19 Nov 2025 GPT-5.1-Codex-Max 58.1 Terminal-Bench 2.0, xhigh
18 Nov 2025 Gemini 3 Pro 54.2 Terminal-Bench 2.0 (Terminus-2 agent)

28 system cards report it, from 17 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 Jun 2026 Ling-2.6-1T technical report 40.5 terminal-bench 2.0, accuracy (Table 5) Capabilities
12 Jun 2026 Muse Spark system card 59 Terminal-Bench 2.0 (2.3.1.1 AI R&D text) AI R&D risk
2 Jun 2026 MAI-Thinking-1 technical report 46 Terminal-Bench 2.0, bash tool only, timeouts ignored (Table 11) Capabilities
2 Jun 2026 MAI-Code-1-Flash model card 54.8 Terminal Bench 2, pass rate (Coding table) Capabilities
26 May 2026 MiniMax-M2.7 technical report 57 Terminal-Bench 2.0, Terminus-2, verified 2.0 dataset (zai-org/terminal-bench-2-verified), 4 trials (Table 4) Capabilities
26 May 2026 MiniMax-M2.5 technical report 51.7 Terminal-Bench 2.0, Terminus-2, verified 2.0 dataset (zai-org/terminal-bench-2-verified), 4 trials (Table 4) Capabilities
25 May 2026 Laguna XS.2 technical report 35.7 Terminal-Bench 2.0, pool harness, 500 steps, mean pass@1 over 4 runs, some task images and verifiers patched (Table 6) Capabilities
25 May 2026 Laguna M.1 technical report 45.8 Terminal-Bench 2.0, pool harness, 500 steps, mean pass@1 over 4 runs, some task images and verifiers patched (Table 7) Capabilities
26 Apr 2026 DeepSeek-V4-Pro technical report 67.9 Terminal Bench 2.0 (Acc), original dataset, DS-V4-Pro Max (Table 6). About 72.0 on the Verified subset (5.3.1 text) Capabilities
26 Apr 2026 DeepSeek-V4-Flash technical report 56.9 Terminal Bench 2.0 (Acc), DeepSeek-V4-Flash Max (Table 7) Capabilities
16 Apr 2026 Claude Opus 4.7 system card 69.4 Terminal-Bench 2.0, Terminus-2 harness, thinking disabled, mean reward over 5 attempts on 89 tasks (Table 8.1.A, 8.3) Capabilities
7 Apr 2026 Claude Mythos Preview system card 82 Terminal-Bench 2.0, Terminus-2 harness in Harbor, adaptive thinking, max effort, 1x timeouts, mean reward over 5 attempts on 89 tasks (Table 6.3.A, 6.5); 92.1 with the 2.1 fixes and 4-hour timeouts Capabilities
25 Mar 2026 Composer 2 technical report 61.7 Terminal-Bench, Cursor's own harness, version not stated, cites the ICLR 2026 Terminal-Bench paper (Table 1 and 7.2) Capabilities
21 Mar 2026 Seed1.8 model card 45.2 Terminal Bench 2.0 (Table 4) Capabilities
19 Mar 2026 Nemotron-Cascade 2 30B-A3B technical report 21.1 Terminal Bench 2.0, Terminus-2, avg@5 (Table 1) Capabilities
11 Mar 2026 Nemotron 3 Super technical report 31 Terminal Bench Core 2.0, Harbor harness (Table 5) Capabilities
28 Feb 2026 qwen3-coder-next technical report 36.2 Terminal-Bench 2.0, Terminus2-json, as in Figure 1. Terminus2-xml 34.2, ClaudeCode 30.9, QwenCode 25.8 (Table 5) Capabilities
19 Feb 2026 Gemini 3.1 Pro model card 68.5 Terminal-Bench 2.0, Terminus-2 harness (results table, p.4) Capabilities
17 Feb 2026 GLM-5 technical report 56.2 Terminal-Bench 2.0 (Terminus-2), original dataset (Table 7). Verified dataset 60.7; Claude Code 56.2 / 61.1 Capabilities
17 Feb 2026 Claude Sonnet 4.6 system card 59.1 Terminal-Bench 2.0, max effort, no thinking budget, Terminus-2 harness (Table 2.1.A, 2.3) Capabilities
14 Feb 2026 Seed2.0 Pro model card 55.8 Terminal Bench 2.0, Terminus 2, 3 tasks excluded (Table 11) Capabilities
11 Feb 2026 Step 3.5 Flash technical report 51 Terminal-Bench 2.0, avg@8 50.98%, pass@8 67/89 (Table 5 and E.2.2) Capabilities
5 Feb 2026 Claude Opus 4.6 system card 65.4 Terminal-Bench 2.0, Terminus-2 harness, adaptive thinking, max effort, 89 tasks x 15 runs (Table 2.3.A, Figure 2.5.A); 61.1 at medium and 55.1 at low effort Capabilities
30 Jan 2026 Kimi K2.5 technical report 50.8 Terminal Bench 2.0, Terminus-2, non-thinking mode (Table 4) Capabilities
6 Jan 2026 MiMo-V2-Flash technical report 38.5 Terminal Bench 2.0 (Table 9) Capabilities
5 Jan 2026 K-EXAONE technical report 29 Terminal-Bench 2.0, Terminus 2 agent (Table 3, Reasoning mode) Capabilities
1 Dec 2025 DeepSeek-V3.2 technical report 46.4 Terminal Bench 2.0 (Acc), Claude Code framework, DeepSeek-V3.2 Thinking (Table 2). 39.3 with Terminus in non-thinking mode Capabilities
24 Nov 2025 Claude Opus 4.5 system card 59.3 Terminal-bench 2.0, Terminus-2 harness, 128k thinking budget (Table 2.3.A; 57.8% with 64k thinking, 2.5) 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 Replaced by Terminal-Bench 2.1

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%