BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases9

Terminal-Bench Hard

Hard subset of Terminal-Bench in a fixed harness. Run by Artificial Analysis.

As of 25 Oct 2026, GPT-5.6 Sol leads Terminal-Bench Hard on BenchLeader with 65.9%, ahead of Claude Fable 5 at 62.9%, across 406 model configurations with a published result.

Published by
Artificial Analysis
Category
Agents & tools
Index weight
0.5
Models
406
Data as of
25 Oct 2026
Newest result
9 Jul 2026

The publisher has posted no new results on this board since 9 Jul 2026. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.

Source: Artificial Analysis (artificialanalysis.ai). Data taken from the public leaderboard.

What the test looks like

Terminal-Bench gives an agent a shell in a container and a task such as building a project from source, recovering a corrupted file system, or setting up a service, then checks the end state with tests.

How it is scored

Percent of tasks passing their checks in that harness.

What to keep in mind

The hardest subset of Terminal-Bench tasks, run by Artificial Analysis in one fixed agent harness for every model. A fixed harness makes models comparable but can understate those built for other agent loops. The subset is small, so scores are noisy. See also the Terminal-Bench board.

Top 15
  1. 1GPT-5.6 Sol (max)65.9%
  2. 2Claude Fable 5 (max)62.9%
  3. 3GPT-5.6 Sol (medium)62.9%
  4. 4GPT-5.6 Terra (xhigh)62.9%
  5. 5GPT-5.6 Sol (high)62.1%
  6. 6GPT-5.6 Sol (xhigh)61.4%
  7. 7GPT-5.5 (xhigh)60.6%
  8. 8GPT-5.6 Sol (low)60.6%
  9. 9GPT-5.5 (high)59.9%
  10. 10Claude Opus 4.8 (max)58.3%
  11. 11GPT-5.4 (xhigh)57.6%
  12. 12GPT-5.6 Terra (max)57.6%
  13. 13GPT-5.5 (medium)57.6%
  14. 14GPT-5.6 Terra (high)57.6%
  15. 15Claude Opus 4.7 (none)54.5%
406 of 406
#
1GPT-5.6 Solmax65.9%68.02026-07-09
2Claude Fable 5max62.9%69.32026-06-09
3GPT-5.6 Solmedium62.9%64.42026-07-09
4GPT-5.6 Terraxhigh62.9%62.82026-07-09
5GPT-5.6 Solhigh62.1%66.32026-07-09
6GPT-5.6 Solxhigh61.4%66.62026-07-09
7GPT-5.5xhigh60.6%66.12026-04-23
8GPT-5.6 Sollow60.6%62.22026-07-09
9GPT-5.5high59.9%65.32026-04-23
10Claude Opus 4.8max58.3%63.32026-05-28
11GPT-5.4xhigh57.6%64.72026-03-05
12GPT-5.6 Terramax57.6%64.42026-07-09
13GPT-5.5medium57.6%62.92026-04-23
14GPT-5.6 Terrahigh57.6%61.32026-07-09
15Claude Opus 4.7no reasoning54.5%58.12026-04-16
16Gemini 3.1 Pro53.8%60.52026-02-19
17GPT-5.3-Codexxhigh53.0%63.82026-02-05
18Claude Sonnet 4.6max53.0%56.02026-02-17
19GPT-5.5low52.3%59.72026-04-23
20GPT-5.4 Minixhigh52.3%56.32026-03-17
21Claude Opus 4.7max51.5%62.42026-04-16
22GLM 5.2max50.8%62.32026-06-16
23Qwen3.7max50.8%60.62026-05-19
24GPT-5.5no reasoning49.2%52.82026-04-23
25KAT-Coder-Pro V249.2%52.42026-03-27
26Claude Opus 4.6no reasoning48.5%56.52026-02-05
27GPT-5.2xhigh47.0%60.72025-12-11
28Claude Opus 4.5thinking47.0%59.92025-11-24
29Qwen3.7 Plus47.0%58.52026-06-01
30DeepSeek V4 Pro 0424max46.2%59.62026-04-24
31Claude Opus 4.6max46.2%57.42026-02-05
32Gemini 3.5 Flashminimal46.2%56.42026-05-19
33Claude Sonnet 4.6no reasoning46.2%54.02026-02-17
34Muse Spark45.5%64.32026-04-08
35GPT-5.1high45.5%57.72025-11-13
36Kimi K2.7 Code44.7%55.82026-06-12
37Qwen3.6 Maxmax43.9%62.32026-04-20
38Kimi K2.6thinking43.9%59.42026-04-20
39Qwen3.6 Plus43.9%58.82026-04-02
40GPT-5.6 Terralow43.9%56.52026-07-09
41GPT-5.4low43.2%59.72026-03-05
42GLM-5thinking43.2%58.92026-02-11
43GPT-5.2medium43.2%57.62025-12-11
44MiMo-V2.5-Prothinking43.2%57.02026-04-22
45GLM 5.1thinking43.2%56.62026-04-07
46GPT-5.5 Instant42.4%58.02026-05-05
47MiniMax M342.4%56.42026-06-01
48GPT-5.4 Nanoxhigh42.4%54.22026-03-17
49Claude Sonnet 4.6low42.4%53.52026-02-17
50Gemini 3 Prohigh41.7%60.02025-11-18
51DeepSeek V4 Pro 0424high41.7%57.82026-04-24
52MiMo-V2.541.7%54.32026-04-22
53Gemini 3.5 Flashhigh40.9%63.12026-05-19
54Mimo v2 Pro40.9%58.62026-03-18
55Qwen3.5 397B A17Bthinking40.9%55.12026-02-16
56Claude Opus 4.5no reasoning40.9%53.62025-11-24
57Gemini 3.5 Flashmedium39.4%62.82026-05-19
58MiniMax M2.739.4%55.62026-03-18
59GLM-5no reasoning39.4%52.32026-02-11
60Gemini 3 Flashthinking38.6%59.12025-12-17

More agents & tools benchmarks · All LLM benchmarks

Cite as: BenchLeader, “Terminal-Bench Hard leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/aa_terminalbench_hard.

Terminal-Bench Hard: questions

What does Terminal-Bench Hard measure?
Terminal-Bench gives an agent a shell in a container and a task such as building a project from source, recovering a corrupted file system, or setting up a service, then checks the end state with tests. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads Terminal-Bench Hard?
GPT-5.6 Sol leads Terminal-Bench Hard with 65.9% as of 25 Oct 2026, ahead of Claude Fable 5 at 62.9%.
How many models have Terminal-Bench Hard results?
406 model configurations have a Terminal-Bench Hard result on BenchLeader, all taken from Artificial Analysis.
Who runs Terminal-Bench Hard and how often is it updated?
Terminal-Bench Hard is published by Artificial Analysis. BenchLeader re-reads the published results once a day and records the date each result was published.
Does Terminal-Bench Hard count toward the BenchLeader Index?
Yes. Terminal-Bench Hard contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.