Claude Sonnet 5.5 records 0.82 on PoetryBench creative coherence metric

The model was evaluated on 1,200 held-out prompts spanning formal verse, constrained rhyme, and open-form composition. Internal logs show median human preference scores of 4.3/5 against 3.1/5 for the prior release. Token-level analysis indicates tighter adherence to meter constraints without increased refusal rate on sensitive themes.

Anthropic's announcement omits direct comparison to GPT-4o or Gemini 1.5 Pro on the same benchmark suite. Independent replication on the public PoetryBench subset confirms the delta but shows smaller margins once prompt distribution is stratified by difficulty. The gain correlates with documented increases in long-context coherence rather than novel architectural changes.

Operationally this shifts internal content pipelines toward native generation of marketing copy and narrative prototypes, cutting external vendor spend by an estimated 30 percent at scale. Deployment telemetry will reveal whether the same capability increase appears in production traffic or remains benchmark-specific.

Next milestone is scheduled for Q2 2026 with expected integration of real-time style transfer from user-provided corpora.