Testing Gemini 4 Carbon on Real Work and Bug Fixes

The author tested the Gemini 4 Carbon model on real work across two repositories with 105 bugs, finding that it scored 57.3 points at a cost of $31.32. This performance places it just behind Fable 5.2, outperforming several other models including Sonnet 5.5 and GPT-6 Astra. The results confirm Google's benchmark claims of significant improvements over Argon at a fraction of competitors' costs. Open article →