Benchmark Almanac

WRENCH (Docker Compose)

Claude Sonnet 5 0.81 leads since September 2026
Not in model cards

Agents maintain a service pipeline through seeded disruptions. The grader measures retained throughput and recovery.

Claude Sonnet 5 and GPT-5.1 tied at 0.98 retained throughput after component loss. Gemini 2.5 Pro reached 0.97. The table gives no run date.

Made by
WRENCH authors
Released
September 2026
Kind
Terminal tasks
Level
University
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Actions
Interaction
Agentic
Environment
Terminal, Docker Compose
Items from
Generated
Refresh
One fixed set
Run by
Anyone
In views
Agents
Strongest model tested
Claude Sonnet 5 #35 on the capability index
Last model added
June 2026
Scores refreshed
November 21st, 2026
3
Models scored on its own leaderboard

Claude Sonnet 5 leads with 0.81 among models released by June 2026

Best score per model from the benchmark's own leaderboard, read on November 21st, 2026. The paper reports no overall score, so this is the plain mean of its four fault-kind throughput-retention scores. Where a model ran with several settings or harnesses, we keep its best.

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

  1. 1 Claude Sonnet 5 0.81
  2. 2 GPT-5.1 0.71
  3. 3 Gemini 2.5 Pro Jun 2025 0.54

Source: Official WRENCH (Docker Compose) leaderboard

#ModelScoreSettingDateSource
1 Claude Sonnet 5 0.81 Jun 2026 Official leaderboard
2 GPT-5.1 0.71 Nov 2025 Official leaderboard
3 Gemini 2.5 Pro Jun 2025 0.54 Jun 2025 Official leaderboard

Rankings by part

We read 4 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as entity destruction, belt cut and resource exhaustion. GPT-5.1 leads 2 of them, and Claude Sonnet 5 leads the other 2. Arrows show how far a model moved from its overall rank among the models we track.

entity destruction

3 models we track have a score here.

  1. 1 GPT-5.1 ↑1 0.98
  2. 2 Claude Sonnet 5 ↓1 0.98
  3. 3 Gemini 2.5 Pro Jun 2025 0.97

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Own leaderboard Not tested since Sep 2026

Source: Official WRENCH (Docker Compose) leaderboard

Scores over time

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

Plot with other benchmarks
On the benchmark’s own leaderboard
0.50 0.60 0.70 0.80 0.90 2025 2026 0.50 0.60 0.70 0.80 0.90 2025 2026

Source: Official WRENCH (Docker Compose) leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Terminal agents

All 35 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 66.5%
  • 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%
  • Terminal-Bench 4.1 In use The Terminal-Bench 4.0 tasks with hardened verifiers: agents complete 66 hard tasks in containerized terminals across software, ML, science, security, and operations. Test scripts verify each result, scored as resolution rate. Copilot Fusion 71.2%
  • 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%
  • Terminal-Bench Pro Not in model cards Agents solve 400 terminal tasks across eight domains. Harbor distributes the 200 public tasks. Alibaba keeps the other 200 for private evaluation. Claude Opus 4.5 51%

Data audit

All 8 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
  • Pass Independent runs Scores from the official leaderboard (3 entries).
  • Note Models scored Only 3 models we track.
  • Pass Newest score The newest is from June 2026, for Claude Sonnet 5.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Claude Sonnet 5 (#35).
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on github.com: 3 rows for models we track at the last read.
  • Pass Last read November 21st, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check