Same model.7% alone, 29% with Edge.
Claude Opus 5.5 on 41 public GitChameleon 2.0 coding tasks it got wrong the first time, share of runs that pass, in our test. Every finished study is below, with what it tested.
3finished studies
85tasks across them
+22 ppwith the Edge connector (GitChameleon 2.0)
Every finished study
| Study | Alone | With Edge or skills | Change | What was tested | Tasks | Date | Report |
|---|---|---|---|---|---|---|---|
| GitChameleon 2.0 | 7% | 29% | +22 pp | Edge connector, with fixes found by an earlier agent (our test) | 41 | Nov 2026 | Report |
| SkillsBench sample | 31% | 54% | +24 pp | Curated skills given directly, agent told to use them (no Edge search) | 32 | Oct 2026 | Study page |
| Personal workflow skills | 17% | 58% | +42 pp | Personal skills given directly, tasks written for the study (exploratory pilot, no Edge search) | 12 | Oct 2026 | Study page |
On one scale
Outlined: the same model alone. Filled: with Edge, or with the skills given directly.
How we test
What does each study test?
GitChameleon 2.0 tests the Edge connector: Claude can call Edge during the task, and Edge brings fixes an earlier agent found (in our test). The two skills studies give the agent the skills directly, with no Edge search, so they measure the skills themselves.
How are runs counted?
Each task runs more than once per side, and every run counts. A run passes when the benchmark’s own tests pass (GitChameleon 2.0, SkillsBench) or every check written for the task passes (personal workflows).
Is it the same model on both sides?
Yes. A study keeps the model and the harness fixed. The only change is the help the agent gets.
How were the tasks chosen?
GitChameleon 2.0: public tasks Claude got wrong the first time. SkillsBench sample: public SkillsBench tasks with curated skills. Personal workflows: tasks written for the study, an exploratory pilot.
Where is the rest?
The method: How we test. Results others published: Published skill evaluations. Outputs side by side: Skill battles.
