Gemini 4 Carbon vs Claude Sonnet 5.5: benchmarks and token cost

This comparison uses Gemini 4 Carbon (High) and Claude Sonnet 5.5 (Max, Default Fallback). It reports the captured benchmark data, with no new hands-on test.

A higher overall score at a higher rate

Carbon’s captured intelligence index is 59.4, compared with Sonnet’s 54.2. Sonnet is listed at $2 input and $10 output per million tokens, against Carbon’s $4 and $20. The useful test is whether Carbon’s aggregate advantage survives the tasks you need, and whether Sonnet’s lower rate changes the cost per accepted answer.

How to read the table

The intelligence index is an aggregate measurement, not a percentage of correct answers in your application. Input and output prices are separate rates per million tokens. Coding and math indices are unavailable in this capture and must not be read as zero. Compare first-token latency separately from output-token throughput, and use the same model effort settings for a trial. Choose using task quality, total time and the cost of accepted results; provider features and deployment requirements need their own checks.