I’m glad this got sorted out, thank you Pawel 🙏
PSA to community, Gemini CLI should 💯 not be used for benchmarks at this point. I speak with some authority on the matter 😅
Testing Gemini Argon 4 on real work with Antigravity CLI, not Gemini CLI, changed the score dramatically.
Argon went from "good" to a frontier model, landing very close to Opus 5.5 (max). I ran the benchmark three times to confirm the results.
It hadn't worked for me before, and I didn't expect the harness to matter so much. The lower cost comes from using fewer tokens and more cache hits.
Thanks for pushing me in the comments! 🙏