LongCLI-Bench: leaderboard
Metric: Pass rate (%), mean of three attempts: share of the 20 long-horizon command-line programming tasks of LongCLI-Bench (CS course assignments and real research and engineering workflows: from scratch, feature addition, bug fix and refactor; isolated Docker environments; about 15,000 lines of code per task); each agent runs through its own harness (Codex, Claude Code or OpenHands) with a unified system prompt; counted passed when both the fail-to-pass requirement tests and the pass-to-pass regression tests reach 100 percent of their steps; higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 16.7 | #60 |
| 2 | Claude Opus 4.5 | 16.7 | #79 |
| 3 | GPT-5.3 Codex | 15 | #68 |
| 4 | GPT-5.1 Codex Max | 11.7 | #94 |
| 5 | Claude Sonnet 4.5 | 10 | #138 |
| 6 | GPT-5.2 Codex | 10 | #89 |
| 7 | Qwen 3 235B A22B | 6.7 | #304 |
| 8 | GLM-4.6 | 6.7 | #246 |
| 9 | DeepSeek V3.1 | 5 | #260 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=longcli-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.