AgentSWE
Harbor: pendingA benchmark of 25 tasks measuring coding agents' ability to build, edit, and optimize other agents against natural-language requirements using held-out cases.
- Samples
- 25
- Released
- 2026
- Scores
- 5
How to run
ltb · browsecomp
$ npm i -g lettokenburn-cli$ ltb bench pull browsecomp --n 50$ ltb bench run ./browsecomp -a claude-code -m anthropic/claude-opus-4-8$ ltb bench push
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 4.6 | 84.0 | N/A | Self-reported | low | Jun 1, 2026 | openai.com |
| GPT-5.4 | 82.7 | N/A | Self-reported | low | Jun 1, 2026 | openai.com |
| Qwen 3.5 | 78.6 | N/A | Self-reported | low | Jun 1, 2026 | openai.com |
| GLM-5 | 75.9 | N/A | Self-reported | low | Jun 1, 2026 | openai.com |
| Claude Sonnet 4.6 | 74.7 | N/A | Self-reported | low | Jun 1, 2026 | openai.com |
| Gemini 3.1 Pro | 59.2 | N/A | Self-reported | low | Jun 1, 2026 | openai.com |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
BrowseComp (Browse Competition) — 网页搜索研究任务. Scoring: Pass@1. Difficulty: 高. Imported from benchmark_research.md (2026-06).