Benchmark Almanac

Terminal-Bench 2.1

GPT-6 Astra 87.4% leads since September 2026
SWE-2 92.8% leads since September 2026
In use

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

Fixed 28 of the 89 tasks in 2.0. Scores moved by up to 12 points, so the two are kept apart.

Released
May 2026
Kind
Terminal tasks
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Strongest model tested
GPT-6 Astra #3 on the capability index
Last in a launch post
November 2026
Last model added
November 2026
Scores refreshed
November 12th, 2026

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)
13
Models scored on its own leaderboard
56
Launch posts list it, from 28 labs
22
System cards report it, from 11 labs

GPT-6 Astra leads with 87.4% among models released by September 2026

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

SWE-2 leads with 92.8%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.

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

13 models. Scroll sideways for the rest.

  1. 87.4 GPT-6 Astra
  2. 83.8 Claude Fable 5
  3. 83.2 GPT-5.5
  4. 79.3 Grok 4.5
  5. 78.9 Claude Opus 4.8
  6. 78.4 GPT-5.6 Terra
  7. 76.2 Muse Spark 1.1
  8. 75.7 GPT-5.6 Luna
  9. 74.6 Claude Sonnet 5
  10. 73.9 Gemini 3 Pro
  11. 68.9 Claude Opus 4.7
  12. 65.8 Gemini 3.1 Pro
  13. 58.6 GLM-5.1
  1. 1 GPT-6 Astra high 87.4%
  2. 2 Claude Fable 5 xhigh 83.8%
  3. 3 GPT-5.5 xhigh 83.2%
  4. 4 Grok 4.5 high 79.3%
  5. 5 Claude Opus 4.8 high 78.9%
  6. 6 GPT-5.6 Terra max 78.4%
  7. 7 Muse Spark 1.1 xhigh 76.2%
  8. 8 GPT-5.6 Luna max 75.7%
  9. 9 Claude Sonnet 5 high 74.6%
  10. 10 Gemini 3 Pro high 73.9%
  11. 11 Claude Opus 4.7 max 68.9%
  12. 12 Gemini 3.1 Pro high 65.8%
  13. 13 GLM-5.1 max 58.6%

Source: Official Terminal-Bench 2.1 leaderboard

62 models. Scroll sideways for the rest.

  1. 92.8 SWE-2
  2. 91.9 GPT-5.6 Sol
  3. 90.6 DeepSeek V4.1 Flash
  4. 89.9 MiMo-V2.6 Pro
  5. 89.4 Gemini 3.8 Flash
  6. 89.2 SWE-2 Flash
  7. 88.8 Muse Spark 1.3
  8. 88.4 LongCat-2.5
  9. 88.3 Kimi K3
  10. 88.2 GLM-5.3
  11. 88.0 Claude Fable 5
  12. 87.9 DeepSeek V4 Pro 0813
  13. 87.4 GPT-6 Astra
  14. 87.4 GPT-5.6 Terra
  15. 86.6 Qwen 3.8 Max
  16. 85.8 Gemini 3.7 Flash
  17. 85.4 Tencent Hy4 preview
  18. 85.0 Step 5 Preview
  19. 84.7 GPT-5.6 Luna
  20. 84.3 GLM-5.3-Flash
  21. 83.9 DeepSeek-V4-Flash-Vision-Exp
  22. 83.3 Grok 4.5
  23. 83.2 GPT-5.5
  24. 82.9 Muse Spark 1.2
  25. 82.7 GLM-5.2
  26. 82.7 DeepSeek V4 Flash 0731
  27. 81.5 SWE-1.7
  28. 81.2 Ling-3.1-flash
  29. 80.5 Ripple-65B Preview
  30. 80.4 Claude Sonnet 5
  31. 80.1 Beam 1.0
  32. 80.0 Muse Spark 1.1
  33. 78.9 Claude Opus 4.8
  34. 78.0 Gemini 3.6 Flash
  35. 76.4 Muse Spark Open
  36. 76.2 Gemini 3.5 Flash
  37. 73.9 Gemini 3 Pro
  38. 73.0 Qwen 3.8 27B
  39. 71.7 Tencent Hy3
  40. 71.0 Seed2.1 Pro
  41. 70.8 LongCat-2.0
  42. 70.2 Laguna S 2.1
  43. 68.9 Claude Opus 4.7
  44. 66.0 MiniMax-M3
  45. 65.8 Gemini 3.1 Pro
  46. 64.7 Inkling-Small
  47. 64.0 Intern-S2-397B
  48. 63.8 Inkling
  49. 60.7 KAT-Coder-Pro V2.5
  50. 60.7 Nemotron 3.5 Super VL
  51. 59.6 Step 3.7 Flash
  52. 58.6 GLM-5.1
  53. 57.0 Solar Pro 4
  54. 56.4 Nemotron 3 Ultra
  55. 54.0 Gemini 3.5 Flash-Lite
  56. 51.7 Muse Glimmer
  57. 51.0 Nemotron 3.5 Bolt 74B-A9B
  58. 43.8 K-EXAONE 2.0
  59. 41.0 KAT-Coder-V2.5-Dev
  60. 29.2 Granite 4.2 30B
  61. 24.6 Nemotron 3.5 Lightning 30B-A3B
  62. 12.9 LFM3-1.7B
  1. 1 SWE-2 92.8%
  2. 2 GPT-5.6 Sol 91.9%
  3. 3 DeepSeek V4.1 Flash 90.6%
  4. 4 MiMo-V2.6 Pro 89.9%
  5. 5 Gemini 3.8 Flash 89.4%
  6. 6 SWE-2 Flash 89.2%
  7. 7 Muse Spark 1.3 88.8%
  8. 8 LongCat-2.5 88.4%
  9. 9 Kimi K3 88.3%
  10. 10 GLM-5.3 88.2%
  11. 11 Claude Fable 5 88%
  12. 12 DeepSeek V4 Pro 0813 87.9%
  13. 13 GPT-6 Astra high 87.4%
  14. 14 GPT-5.6 Terra 87.4%
  15. 15 Qwen 3.8 Max 86.6%
  16. 16 Gemini 3.7 Flash 85.8%
  17. 17 Tencent Hy4 preview 85.4%
  18. 18 Step 5 Preview 85%
  19. 19 GPT-5.6 Luna 84.7%
  20. 20 GLM-5.3-Flash 84.3%
  21. 21 DeepSeek-V4-Flash-Vision-Exp 83.9%
  22. 22 Grok 4.5 83.3%
  23. 23 GPT-5.5 xhigh 83.2%
  24. 24 Muse Spark 1.2 82.9%
  25. 25 GLM-5.2 82.7%
  26. 26 DeepSeek V4 Flash 0731 82.7%
  27. 27 SWE-1.7 81.5%
  28. 28 Ling-3.1-flash 81.2%
  29. 29 Ripple-65B Preview 80.5%
  30. 30 Claude Sonnet 5 80.4%
  31. 31 Beam 1.0 80.1%
  32. 32 Muse Spark 1.1 80%
  33. 33 Claude Opus 4.8 high 78.9%
  34. 34 Gemini 3.6 Flash 78%
  35. 35 Muse Spark Open 76.4%
  36. 36 Gemini 3.5 Flash 76.2%
  37. 37 Gemini 3 Pro high 73.9%
  38. 38 Qwen 3.8 27B 73%
  39. 39 Tencent Hy3 71.7%
  40. 40 Seed2.1 Pro 71%
  41. 41 LongCat-2.0 70.8%
  42. 42 Laguna S 2.1 70.2%
  43. 43 Claude Opus 4.7 max 68.9%
  44. 44 MiniMax-M3 66%
  45. 45 Gemini 3.1 Pro high 65.8%
  46. 46 Inkling-Small 64.7%
  47. 47 Intern-S2-397B 64%
  48. 48 Inkling 63.8%
  49. 49 KAT-Coder-Pro V2.5 60.7%
  50. 50 Nemotron 3.5 Super VL 60.7%
  51. 51 Step 3.7 Flash 59.6%
  52. 52 GLM-5.1 max 58.6%
  53. 53 Solar Pro 4 57%
  54. 54 Nemotron 3 Ultra 56.4%
  55. 55 Gemini 3.5 Flash-Lite 54%
  56. 56 Muse Glimmer 51.7%
  57. 57 Nemotron 3.5 Bolt 74B-A9B 51%
  58. 58 K-EXAONE 2.0 43.8%
  59. 59 KAT-Coder-V2.5-Dev 41%
  60. 60 Granite 4.2 30B 29.2%
  61. 61 Nemotron 3.5 Lightning 30B-A3B 24.6%
  62. 62 LFM3-1.7B 12.9%

Sources: Labs’ model cards and official Terminal-Bench 2.1 leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 87.4% high, Codex Sep 2026 Official leaderboard
2 Claude Fable 5 83.8% xhigh, Claude Code Jun 2026 Official leaderboard
3 GPT-5.5 83.2% xhigh, Codex Apr 2026 Official leaderboard
4 Grok 4.5 79.3% high, Cursor CLI Jul 2026 Official leaderboard
5 Claude Opus 4.8 78.9% high, Claude Code May 2026 Official leaderboard
6 GPT-5.6 Terra 78.4% max, Codex Jul 2026 Official leaderboard
7 Muse Spark 1.1 76.2% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
8 GPT-5.6 Luna 75.7% max, Codex Jul 2026 Official leaderboard
9 Claude Sonnet 5 74.6% high, Claude Code Jun 2026 Official leaderboard
10 Gemini 3 Pro 73.9% high, Terminus 2 Nov 2025 Official leaderboard
11 Claude Opus 4.7 68.9% max, Claude Code Apr 2026 Official leaderboard
12 Gemini 3.1 Pro 65.8% high, Gemini CLI Feb 2026 Official leaderboard
13 GLM-5.1 58.6% max, Claude Code Apr 2026 Official leaderboard
#ModelScoreSettingDateSource
1 SWE-2 92.8% Sep 2026 Lab model card
2 GPT-5.6 Sol 91.9% Jul 2026 Lab model card
3 DeepSeek V4.1 Flash 90.6% Sep 2026 Lab model card
4 MiMo-V2.6 Pro 89.9% Sep 2026 Lab model card
5 Gemini 3.8 Flash 89.4% Sep 2026 Lab model card
6 SWE-2 Flash 89.2% Nov 2026 Lab model card
7 Muse Spark 1.3 88.8% Sep 2026 Lab model card
8 LongCat-2.5 88.4% Nov 2026 Lab model card
9 Kimi K3 88.3% Jul 2026 Lab model card
10 GLM-5.3 88.2% Aug 2026 Lab model card
11 Claude Fable 5 88% Jun 2026 Lab model card
12 DeepSeek V4 Pro 0813 87.9% Aug 2026 Lab model card
13 GPT-6 Astra 87.4% high, Codex Sep 2026 Official leaderboard
14 GPT-5.6 Terra 87.4% Jul 2026 Lab model card
15 Qwen 3.8 Max 86.6% Aug 2026 Lab model card
16 Gemini 3.7 Flash 85.8% Aug 2026 Lab model card
17 Tencent Hy4 preview 85.4% Aug 2026 Lab model card
18 Step 5 Preview 85% Sep 2026 Lab model card
19 GPT-5.6 Luna 84.7% Jul 2026 Lab model card
20 GLM-5.3-Flash 84.3% Aug 2026 Lab model card
21 DeepSeek-V4-Flash-Vision-Exp 83.9% Aug 2026 Lab model card
22 Grok 4.5 83.3% Jul 2026 Lab model card
23 GPT-5.5 83.2% xhigh, Codex Apr 2026 Official leaderboard
24 Muse Spark 1.2 82.9% Aug 2026 Lab model card
25 GLM-5.2 82.7% Jun 2026 Lab model card
26 DeepSeek V4 Flash 0731 82.7% Jul 2026 Lab model card
27 SWE-1.7 81.5% Jul 2026 Lab model card
28 Ling-3.1-flash 81.2% Oct 2026 Lab model card
29 Ripple-65B Preview 80.5% Nov 2026 Lab model card
30 Claude Sonnet 5 80.4% Jun 2026 Lab model card
31 Beam 1.0 80.1% Oct 2026 Lab model card
32 Muse Spark 1.1 80% Jul 2026 Lab model card
33 Claude Opus 4.8 78.9% high, Claude Code May 2026 Official leaderboard
34 Gemini 3.6 Flash 78% Jul 2026 Lab model card
35 Muse Spark Open 76.4% Oct 2026 Lab model card
36 Gemini 3.5 Flash 76.2% May 2026 Lab model card
37 Gemini 3 Pro 73.9% high, Terminus 2 Nov 2025 Official leaderboard
38 Qwen 3.8 27B 73% Aug 2026 Lab model card
39 Tencent Hy3 71.7% Jul 2026 Lab model card
40 Seed2.1 Pro 71% Jun 2026 Lab model card
41 LongCat-2.0 70.8% Jul 2026 Lab model card
42 Laguna S 2.1 70.2% Jul 2026 Lab model card
43 Claude Opus 4.7 68.9% max, Claude Code Apr 2026 Official leaderboard
44 MiniMax-M3 66% Jun 2026 Lab model card
45 Gemini 3.1 Pro 65.8% high, Gemini CLI Feb 2026 Official leaderboard
46 Inkling-Small 64.7% Jul 2026 Lab model card
47 Intern-S2-397B 64% Sep 2026 Lab model card
48 Inkling 63.8% Jul 2026 Lab model card
49 KAT-Coder-Pro V2.5 60.7% Jul 2026 Lab model card
50 Nemotron 3.5 Super VL 60.7% Oct 2026 Lab model card
51 Step 3.7 Flash 59.6% May 2026 Lab model card
52 GLM-5.1 58.6% max, Claude Code Apr 2026 Official leaderboard
53 Solar Pro 4 57% Aug 2026 Lab model card
54 Nemotron 3 Ultra 56.4% Jun 2026 Lab model card
55 Gemini 3.5 Flash-Lite 54% Jul 2026 Lab model card
56 Muse Glimmer 51.7% Aug 2026 Lab model card
57 Nemotron 3.5 Bolt 74B-A9B 51% Oct 2026 Lab model card
58 K-EXAONE 2.0 43.8% Jul 2026 Lab model card
59 KAT-Coder-V2.5-Dev 41% Jul 2026 Lab model card
60 Granite 4.2 30B 29.2% Aug 2026 Lab model card
61 Nemotron 3.5 Lightning 30B-A3B 24.6% Aug 2026 Lab model card
62 LFM3-1.7B 12.9% Oct 2026 Lab model card

Labs report scores for 56 models in their cards. Switch the scores above to rank them together.

Pareto: accuracy against time

Each point is one agent, model and effort. The board plots total run time; this plots the mean time per trial, the same frontier at a fixed trial count. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

60% 70% 80% 90% 100% 1001,00010K Mean time per trial (s), log scale, lower is better Accuracy Gemini 3 Pro (high): Accuracy 73.9%, mean time per trial 548 s Gemini 3 Pro (high): Accuracy 65.8%, mean time per trial 437 s Gemini 3.1 Pro (high): Accuracy 65.8%, mean time per trial 653 s Gemini 3.1 Pro (high): Accuracy 65.6%, mean time per trial 673 s GLM-5.1 (max): Accuracy 58.6%, mean time per trial 945 s Claude Opus 4.7 (max): Accuracy 68.9%, mean time per trial 800 s Claude Opus 4.7 (max): Accuracy 66.1%, mean time per trial 1,043 s GPT-5.5 (xhigh): Accuracy 83.2%, mean time per trial 483 s GPT-5.5 (xhigh): Accuracy 78%, mean time per trial 727 s Claude Opus 4.8 (high): Accuracy 78.9%, mean time per trial 551 s Claude Fable 5 (xhigh): Accuracy 83.8%, mean time per trial 580 s Claude Fable 5 (high): Accuracy 80.5%, mean time per trial 697 s Claude Sonnet 5 (high): Accuracy 74.6%, mean time per trial 630 s Grok 4.5 (high): Accuracy 79.3%, mean time per trial 443 s GPT-5.6 Luna (max): Accuracy 75.7%, mean time per trial 457 s GPT-5.6 Terra (max): Accuracy 78.4%, mean time per trial 433 s Muse Spark 1.1 (xhigh): Accuracy 76.2%, mean time per trial 687 s GPT-6 Astra (max): Accuracy 86.7%, mean time per trial 576 s GPT-6 Astra (xhigh): Accuracy 85.8%, mean time per trial 487 s GPT-6 Astra (high): Accuracy 87.4%, mean time per trial 434 s GPT-6 Astra (medium): Accuracy 87%, mean time per trial 390 s GPT-6 Astra (low): Accuracy 86.7%, mean time per trial 349 s GPT-6 Astra (high)GPT-6 Astra (medium)GPT-6 Astra (low) 60% 70% 80% 90% 100% 1001,00010K Mean time per trial (s), log scale, lower is better Accuracy Gemini 3 Pro (high): Accuracy 73.9%, mean time per trial 548 s Gemini 3 Pro (high): Accuracy 65.8%, mean time per trial 437 s Gemini 3.1 Pro (high): Accuracy 65.8%, mean time per trial 653 s Gemini 3.1 Pro (high): Accuracy 65.6%, mean time per trial 673 s GLM-5.1 (max): Accuracy 58.6%, mean time per trial 945 s Claude Opus 4.7 (max): Accuracy 68.9%, mean time per trial 800 s Claude Opus 4.7 (max): Accuracy 66.1%, mean time per trial 1,043 s GPT-5.5 (xhigh): Accuracy 83.2%, mean time per trial 483 s GPT-5.5 (xhigh): Accuracy 78%, mean time per trial 727 s Claude Opus 4.8 (high): Accuracy 78.9%, mean time per trial 551 s Claude Fable 5 (xhigh): Accuracy 83.8%, mean time per trial 580 s Claude Fable 5 (high): Accuracy 80.5%, mean time per trial 697 s Claude Sonnet 5 (high): Accuracy 74.6%, mean time per trial 630 s Grok 4.5 (high): Accuracy 79.3%, mean time per trial 443 s GPT-5.6 Luna (max): Accuracy 75.7%, mean time per trial 457 s GPT-5.6 Terra (max): Accuracy 78.4%, mean time per trial 433 s Muse Spark 1.1 (xhigh): Accuracy 76.2%, mean time per trial 687 s GPT-6 Astra (max): Accuracy 86.7%, mean time per trial 576 s GPT-6 Astra (xhigh): Accuracy 85.8%, mean time per trial 487 s GPT-6 Astra (high): Accuracy 87.4%, mean time per trial 434 s GPT-6 Astra (medium): Accuracy 87%, mean time per trial 390 s GPT-6 Astra (low): Accuracy 86.7%, mean time per trial 349 s GPT-6 Astra (high)GPT-6 Astra (medium)GPT-6 Astra (low)
ModelSettingAccuracyMean time per trial
GPT-6 Astra high, Codex 87.4% 434.1 s
GPT-6 Astra medium, Codex 87% 390.4 s
GPT-6 Astra low, Codex 86.7% 348.9 s
GPT-6 Astra max, Codex 86.7% 576 s
GPT-6 Astra xhigh, Codex 85.8% 487.3 s
Claude Fable 5 xhigh, Claude Code 83.8% 579.7 s
GPT-5.5 xhigh, Codex 83.2% 482.6 s
Claude Fable 5 high, Terminus 2 80.5% 697.1 s
Grok 4.5 high, Cursor CLI 79.3% 443 s
Claude Opus 4.8 high, Claude Code 78.9% 551.4 s
GPT-5.6 Terra max, Codex 78.4% 432.8 s
GPT-5.5 xhigh, Terminus 2 78% 726.7 s
Muse Spark 1.1 xhigh, mini-SWE-agent 76.2% 687 s
GPT-5.6 Luna max, Codex 75.7% 457.3 s
Claude Sonnet 5 high, Claude Code 74.6% 629.6 s
Gemini 3 Pro high, Terminus 2 73.9% 547.8 s
Claude Opus 4.7 max, Claude Code 68.9% 799.6 s
Claude Opus 4.7 max, Terminus 2 66.1% 1,043 s
Gemini 3 Pro high, Gemini CLI 65.8% 436.6 s
Gemini 3.1 Pro high, Gemini CLI 65.8% 652.9 s
Gemini 3.1 Pro high, Terminus 2 65.6% 672.8 s
GLM-5.1 max, Claude Code 58.7% 944.6 s

The best score, 87.4%, costs $1.74 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Anthropic xAI OpenAI Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.10 $0.20 $0.50 $1.00 $2.00 $5.00 $10 Cost per task, log scale GPT-6 Astra, high GPT-6 Astra, low Claude Fable 5, high Grok 4.5, high 0% 20% 40% 60% 80% 100% $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra, high Claude Fable 5, high

Source: Official Terminal-Bench 2.1 leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra high Codex 87.4% $1.74 9,485
GPT-6 Astra medium Codex 87% $1.39 6,517
GPT-6 Astra low Codex 86.7% $1.04 4,400
GPT-6 Astra max Codex 86.7% $2.40 17,028
GPT-6 Astra xhigh Codex 85.8% $2.00 11,882
Claude Fable 5 xhigh Claude Code 83.8% $1.24 22,371
GPT-5.5 xhigh Codex 83.2% $4.63 13,408
Claude Fable 5 high Terminus 2 80.5% $0.99 16,657
Grok 4.5 high Cursor CLI 79.3% $0.30 10,638
Claude Opus 4.8 high Claude Code 78.9% $0.64 18,178
GPT-5.6 Terra max Codex 78.4% $0.95 19,567
GPT-5.5 xhigh Terminus 2 78% $1.11 24,198
Muse Spark 1.1 xhigh mini-SWE-agent 76.2% $0.45 30,741
GPT-5.6 Luna max Codex 75.7% $0.54 23,885
Claude Sonnet 5 high Claude Code 74.6% $0.65 25,205
Gemini 3 Pro high Terminus 2 73.9% $0.50 28,905
Claude Opus 4.7 max Claude Code 68.9% $1.34 33,326
Claude Opus 4.7 max Terminus 2 66.1% $1.31 27,980
Gemini 3 Pro high Gemini CLI 65.8% $0.56 16,223
Gemini 3.1 Pro high Gemini CLI 65.8% $0.53 13,761
Gemini 3.1 Pro high Terminus 2 65.6% $0.52 29,069
GLM-5.1 max Claude Code 58.6% $0.62 23,900

3 models have held number one on the official leaderboard since its release in May 2026 4 models have held number one with claims included since its release in May 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.

OpenAI Anthropic Google DeepMind Cognition
Own leaderboard Passed 85% in Sep 2026, by GPT-6 Astra
GPT-6 Astra 87.4%
Lab claims Passed 85% in Jun 2026, by Claude Fable 5
SWE-2 92.8%

Sources: Official Terminal-Bench 2.1 leaderboard and labs’ model cards

Anthropic’s own number for Claude Fable 5 passed 85% in June 2026

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%) SWE-2 Claude Fable 5 GPT-5.5 Gemini 3 Pro 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) SWE-2 GPT-5.5

Sources: Official Terminal-Bench 2.1 leaderboard and labs’ model cards

28 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
Anthropic
StepFun
MiniMax
NVIDIA
Z.ai
ByteDance Seed
Tencent
Meituan
Kuaishou
Cognition
OpenAI
Meta
Thinking Machines
xAI
Moonshot AI
Poolside
DeepSeek
LG AI Research
Qwen
Upstage
IBM
Shanghai AI Lab
Xiaomi Corp
Reflection
Ant Group
Liquid AI
Ricursive Intelligence
DateLaunch postScoreAs the post writes it
Nov 4, 2026 Ripple-65B Preview 80.5 TerminalBench 2.1
Nov 4, 2026 SWE-2 Flash 89.2 Terminal-Bench 2.1
Nov 2, 2026 LongCat-2.5 88.4 Terminal Bench v2.1
Oct 27, 2026 LFM3-1.7B 12.9 Terminal-Bench 2.1 (OpenCode)
Oct 15, 2026 Muse Spark Open 76.4 Terminal-Bench 2.1 Terminus 2 agent
Oct 15, 2026 Nemotron 3.5 Bolt 74B-A9B 51 Terminal-Bench 2.1
Oct 15, 2026 Nemotron 3.5 Super VL 60.7 Terminal-Bench 2.1 (Terminus 2)
Oct 13, 2026 Ling-3.1-flash 81.2 Terminal-Bench-2.1
Oct 5, 2026 Beam 1.0 80.1 Terminal Bench v2.1
Sep 22, 2026 MiMo-V2.6 Pro 89.9 Terminal Bench 2.1
Sep 20, 2026 Step 5 Preview 85 Terminal-Bench v2.1
Sep 13, 2026 Intern-S2-397B 64 TerminalBench 2.1 (Terminus 2)
Sep 10, 2026 SWE-2 92.8 Terminal-Bench 2.1
Sep 10, 2026 DeepSeek V4.1 Flash 90.6 Terminal-Bench 2.1 (Pass@1)
Sep 2, 2026 Muse Spark 1.3 88.8 Terminal-Bench 2.1 (Agentic terminal coding)
Sep 2, 2026 Gemini 3.8 Flash 89.4 Terminal-bench 2.1
Aug 28, 2026 Tencent Hy4 preview 85.4 Terminal-Bench 2.1
Aug 26, 2026 GLM-5.3-Flash 84.3 Terminal Bench 2.1 (Claude Code 2.1.207)
Aug 25, 2026 Granite 4.2 30B 29.2 Terminal-Bench 2.1
Aug 21, 2026 DeepSeek-V4-Flash-Vision-Exp 83.9 Terminal Bench 2.1
Aug 14, 2026 Qwen 3.8 27B 73 Agentic terminal coding Terminal Bench 2.1 (Terminus)
Aug 14, 2026 GLM-5.3 88.2 Terminal Bench 2.1
Aug 13, 2026 DeepSeek V4 Pro 0813 87.9 Terminal Bench 2.1
Aug 13, 2026 Gemini 3.7 Flash 85.8 Terminal-bench 2.1
Aug 11, 2026 Nemotron 3.5 Lightning 30B-A3B 24.6 Terminal-Bench 2.1
Aug 11, 2026 Solar Pro 4 57 Terminal-Bench v2.1 (Artificial Analysis score quoted by Upstage)
Aug 10, 2026 Muse Glimmer 51.7 TerminalBench 2.1 (with terminus2)
Aug 5, 2026 Muse Spark 1.2 82.9 Terminal-Bench 2.1 (Muse Code harness)
Aug 3, 2026 Qwen 3.8 Max 86.6 Terminal Bench 2.1
Jul 31, 2026 K-EXAONE 2.0 43.8 Terminal-Bench 2.1
Jul 31, 2026 DeepSeek V4 Flash 0731 82.7 Terminal Bench 2.1
Jul 30, 2026 Inkling-Small 64.7 Terminal Bench 2.1 (best harness)
Jul 23, 2026 KAT-Coder-V2.5-Dev 41 Terminal-Bench 2.1, mean of Terminus-2 32.60 / Claude Code 49.44
Jul 21, 2026 Laguna S 2.1 70.2 Terminal-Bench 2.1
Jul 21, 2026 Gemini 3.5 Flash-Lite 54 Terminal-bench 2.1 (Terminus-2 harness)
Jul 21, 2026 Gemini 3.6 Flash 78 Terminal-bench 2.1 (Terminus-2 harness)
Jul 16, 2026 Kimi K3 88.3 Terminal-Bench 2.1
Jul 16, 2026 Grok 4.5 83.3 Terminal Bench 2.1
Jul 15, 2026 Inkling 63.8 Terminal Bench 2.1 (best harness)
Jul 9, 2026 GPT-5.6 Luna 84.7 Terminal-Bench 2.1
Jul 9, 2026 GPT-5.6 Terra 87.4 Terminal-Bench 2.1
Jul 9, 2026 Muse Spark 1.1 80 Terminal-Bench 2.1 (Agentic terminal coding)
Jul 9, 2026 GPT-5.6 Sol 91.9 Terminal-Bench 2.1 (Sol Ultra, 4 agents)
Jul 8, 2026 SWE-1.7 81.5 Terminal-Bench 2.1 (Devin CLI, timeout 4h)
Jul 6, 2026 KAT-Coder-Pro V2.5 60.7 Terminal-Bench 2.1 (AA Coding Index category)
Jul 5, 2026 LongCat-2.0 70.8 Terminal-Bench 2.1
Jul 2, 2026 Tencent Hy3 71.7 Terminal-Bench 2.1
Jun 30, 2026 Claude Sonnet 5 80.4 Agentic coding Terminal-Bench 2.1
Jun 23, 2026 Seed2.1 Pro 71 Terminal Bench 2.1
Jun 16, 2026 GLM-5.2 82.7 Terminal Bench 2.1 (Best Reported Harness)
Jun 9, 2026 Claude Fable 5 88 Agentic coding Terminal-Bench 2.1
Jun 4, 2026 Nemotron 3 Ultra 56.4 Terminal Bench 2.1
Jun 1, 2026 MiniMax-M3 66 Terminal Bench 2.1
May 29, 2026 Step 3.7 Flash 59.6 Terminal-Bench 2.1
May 28, 2026 Claude Opus 4.8 74.6 Agentic terminal coding Terminal-Bench 2.1
May 19, 2026 Gemini 3.5 Flash 76.2 Terminal-bench 2.1 (Terminus-2 harness)

22 system cards report it, from 11 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
Oct 19, 2026 Muse Glimmer technical report 51.7 TerminalBench 2.1 (Terminus2), high, 54.3 xhigh (Tables 2 and 5) Capabilities
Oct 15, 2026 Nemotron 3.5 Super VL technical report 60.7 Terminal-Bench 2.1, Terminus 2 harness, 3 hour timeout, mean of 8 runs (Table 10) Capabilities
Oct 15, 2026 Nemotron 3.5 Lightning 30B-A3B technical report 24.4 Terminal-Bench v2.1, BF16 acc. (Table 22) Capabilities
Oct 15, 2026 Nemotron 3.5 Bolt 74B-A9B technical report 51 Terminal-Bench 2.1, released 24GB checkpoint (Table 7). BF16 53.93 Capabilities
Sep 22, 2026 MiMo-V2.6 Pro technical report 89.9 Terminal Bench 2.1 (Table 3) Capabilities
Sep 10, 2026 DeepSeek V4.1 Flash technical report 90.6 Terminal-Bench 2.1 (Pass@1), DeepSeek Harness Minimal (Table 3, Table 4) Capabilities
Sep 2, 2026 Gemini 3.8 Flash model card 89.4 Terminal-bench 2.1 (results table, p.5) Capabilities
Aug 13, 2026 Gemini 3.7 Flash model card 85.8 Terminal-bench 2.1 (results table, p.5) Capabilities
Aug 5, 2026 K-EXAONE 2.0 technical report 43.8 Terminal-Bench 2.1 (Table 6, Reasoning mode) Capabilities
Jul 30, 2026 Inkling-Small model card 64.7 Terminal Bench 2.1, best harness (5. Evaluations table) Capabilities
Jul 27, 2026 Kimi K3 technical report 88.3 Terminal-Bench 2.1, best score across harnesses (Table 2) Capabilities
Jul 21, 2026 Gemini 3.6 Flash model card 78 Terminal-bench 2.1, Terminus-2 harness (results table, p.4) Capabilities
Jul 21, 2026 Gemini 3.5 Flash-Lite model card 54 Terminal-bench 2.1, Terminus-2 harness (results table, p.4) Capabilities
Jul 15, 2026 Inkling model card 63.8 Terminal Bench 2.1, best harness, contaminated rollouts scored 0 (5. Evaluations table, effort 0.99) Capabilities
Jul 14, 2026 Grok 4.5 model card 83.3 Terminal-Bench 2.1, task success rate, high, Grok Build harness (2.9) Capabilities
Jul 9, 2026 Muse Spark 1.1 system card 80 Terminal-Bench 2.1 (Figure 44) Capabilities
Jun 30, 2026 Claude Sonnet 5 system card 80.4 Terminal-Bench 2.1, mean reward over 5 attempts on 89 tasks at xhigh effort, mini-SWE-agent harness (Table 8.1.A, 8.3) Capabilities
Jun 23, 2026 Seed2.1 Pro model card 71 Terminal-Bench 2.1 (Table 5) Capabilities
Jun 9, 2026 Nemotron 3 Ultra technical report 56.4 Terminal Bench 2.1 (Table 10) Capabilities
Jun 9, 2026 Claude Fable 5 system card 84.3 Terminal-Bench 2.1, mini-SWE-agent harness, mean reward over 5 attempts on 89 tasks, high effort, 20.9% of trials hit a safety refusal and fell back to Opus 4.8 (Table 8.1.A, 8.3) Capabilities
May 28, 2026 Claude Opus 4.8 system card 74.6 Terminal-Bench 2.1 mean reward on Harbor with the Terminus-2 harness, high effort, 5 attempts on each of 89 tasks (Table 8.1.A, 8.3) Capabilities
May 19, 2026 Gemini 3.5 Flash model card 76.2 Terminal-bench 2.1, Terminus-2 harness (results table, p.4) Capabilities

The lab's number comes out ahead in 6 of 7 pairs that differ

Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. Every benchmark.

Model Lab says Board says Gap
GPT-5.6 Terra 87.4% Terminal-Bench 2.1 78.4% Leaderboard, max, Codex +9.0 pts
GPT-5.6 Luna 84.7% Terminal-Bench 2.1 75.7% Leaderboard, max, Codex +9.0 pts
Claude Sonnet 5 80.4% Agentic coding Terminal-Bench 2.1 74.6% Leaderboard, high, Claude Code +5.8 pts
Claude Opus 4.8 74.6% Agentic terminal coding Terminal-Bench 2.1 78.9% Leaderboard, high, Claude Code −4.3 pts
Claude Fable 5 88% Agentic coding Terminal-Bench 2.1 83.8% Leaderboard, xhigh, Claude Code +4.2 pts
Grok 4.5 83.3% Terminal Bench 2.1 79.3% Leaderboard, high, Cursor CLI +4.0 pts
Muse Spark 1.1 80% Terminal-Bench 2.1 (Agentic terminal coding) 76.2% Leaderboard, xhigh, mini-SWE-agent +3.8 pts

2 logged changes that move scores

Changes by the maintainers, so a score from before a change may not compare with one from after it. Newest first, each with the maintainers' own source. The list may be incomplete. What we log.

  1. Leaderboard rules changed July 13th, 2026

    The maintainers required ATIF trajectories for rewarded trials during submission checks. Submissions without those trajectories now fail intake despite reporting successful tasks. Source

  2. Grader changed July 11th, 2026

    The maintainers changed the reward-hacking judge to assess task intent instead of gold-file heuristics. The revised rules can change which successful trials the leaderboard disqualifies. Source

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.

Google DeepMind Anthropic OpenAI
Terminal-Bench 4.1 20 leaders
Claude Opus 5.5 65.8%
Terminal-Bench Replaced by 2.0, Nov 2025
GPT-5 49.6%
Terminal-Bench 2.0 Replaced by 2.1, May 2026
GPT-5.5 84.7%
Terminal-Bench 2.1 Passed 85% in Sep 2026, by GPT-6 Astra
GPT-6 Astra 87.4%
Terminal-Bench 3.0 Replaced by 4.0, Aug 2026
Claude Opus 5 42.7%
Terminal-Bench 4.0 4 leaders
Claude Opus 5.5 64.8%
Terminal-Bench 4.1 1 leader
Claude Opus 5.5 65.8%

Sources: Official leaderboards and boards copied by Epoch AI (CC BY 4.0)

Replaces Terminal-Bench 2.0 Replaced by Terminal-Bench 3.0

More in Terminal agents

All 19 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%
  • Terminal-Bench 2.1 Extended Not in model cards Agents complete long command-line tasks in containers, scored by verification tests. Mercor adds 99 new expert-written tasks to the 89 public ones and reports both scores. Muse Wonder 54%
  • o11y-bench Not in model cards Agents query metrics, logs and traces in a live Grafana stack, investigate incidents and fix dashboards. Grading checks claims against facts read from the same environment. Claude Opus 5.5 93.8%
  • 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%