Claude Sonnet 5.5 vs Claude Fable 5.1: benchmarks and token cost

This comparison uses Claude Sonnet 5.5 (Max, Default Fallback) and Claude Fable 5.1 (Xhigh, Default Fallback). It reports the captured benchmark data, with no new hands-on test.

Sonnet and Fable in an Anthropic evaluation

Sonnet’s captured intelligence index is 54.2 and Fable’s is 49.7. Sonnet’s input and output rates are one fifth of Fable’s in this snapshot. That is a clear reason to evaluate Sonnet as an alternative, while retaining the exact effort and fallback labels. A successful migration still needs task-level quality checks and an examination of output and latency behavior.

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.