Gemini 4 Carbon Nears FrontierCode SOTA

Carbon’s result suggests Google has closed much of the gap in agentic coding, but frontier leadership still depends on reliability across messy, real-world repository work.

• –FrontierCode 2.0 evaluates correctness, scope discipline, test quality, regressions, and maintainability across 132 tasks.
• –Carbon’s near-40% score is competitive with the top tier, though Claude Opus 5.5 and GPT-6 Astra remain ahead.
• –The roughly $4.04 task cost makes Carbon attractive for high-volume coding agents.
• –Cognition’s emphasis on unstated requirements and thoroughness makes this more meaningful than benchmarks built around fully specified coding prompts.
• –The result reinforces that coding-model quality is increasingly about autonomous judgment, not just code generation.