Gemini 4 Argon in Cursor vs Gemini CLI: same 7/9 defects fixed, different behavior
On the same repository with 9 planted defects, Gemini 4 Argon fixed 7/9 in both Cursor and Gemini CLI, but Cursor thought longer before acting while the CLI moved faster, at €0.92 in API cost.
Gemini 4 Argon: Cursor vs Gemini CLI (API key) After testing Opus 5.5, I ran Gemini 4 Argon on the exact same repository, with the same 9 planted defects and open-ended coding challenge. Results: > Argon in Cursor 7/9 hidden defect signatures changed 60 tests passing (+7) ~11 minutes total Spent nearly 4 minutes thinking before taking its first action > Argon via Gemini CLI + API key 7/9 hidden defect signatures changed 58 tests passing (+5) ~9 minutes total €0.92 API cost Started taking action much sooner Both preserved the GT-10 restraint condition and declared the repository ready to ship. What caught my attention wasn't just the identical score, but the difference in agent behavior. In Cursor, Argon spent a significant amount of time thinking before touching the repository. Through the API, it moved into execution much faster. Same model. Same challenge. Different harness. Different behavior. Same 7/9 provisional result. A reminder that the coding environment matters just as much as the model we're evaluating. Has anyone else noticed this difference between Cursor and direct API usage?