Opus 5.5 fixes quality, but the effort trap costs more than tokens.

The Model. Anthropic dropped Opus 5.5, ten months after the revolutionary Opus 4.5 and a disappointing Opus 5. The key claim is that it performs at Fable 5.1 levels on most work at 40% lower cost. Input and output token prices dropped to $4 and $20 per million tokens respectively, with cache reads 60% cheaper at $0.20 per million. This pricing positions it well for businesses that require zero data retention policies—something the Fable line does not offer.

The Benchmarks. The model leads the Artificial Analysis intelligence index with a 5-point jump over Fable 5.1. It dominates real-world coding benchmarks like Cursor Bench and Terminal Bench on medium and high effort levels. However, the model is token-hungry: on max effort it averages 120,000 tokens per task versus Fable's 78,000 and GPT Astra's 27,000. The cost advantage comes entirely from price cuts and cache optimizations, not efficiency.

The Effort Trap. The most significant operational insight is that max reasoning level is a trap. The creator found a 10x token jump from X-high to max effort with barely 1% score improvement. He lost a 6.5-hour session to infinite loops on max. The practical advice is to never use max effort on this model. Medium effort delivers Fable 5.1-level intelligence at $1.34 per task versus Fable's $6+. This is the kind of cost-saving behavior change teams should immediately enforce through default configurations.

The Communication Fix. Opus 5.5's writing is dramatically clearer than Opus 5's opaque, dash-heavy prose. One example shows a confusing bug description rewritten into a precise root cause analysis. This matters more than it sounds: developers spend hours reading model output. Clearer communication means faster review cycles and fewer misinterpreted PRs. Early testers report deleting their system prompt unslop configurations. The model also now allows changing reasoning levels mid-session without busting cache, a practical improvement for long coding sessions.

The Weaknesses. The model has quirks that benchmarks miss. It shows paranoia about context window compaction, making nonsensical statements about crashes risking lost work—behavior Fable doesn't exhibit. On a custom code review benchmark measuring depth and accuracy, Opus 5.5 nearly doubled Opus 5's score, but still trailed frontier OpenAI models. Frontend design quality is mixed: the creator found Fable superior for taste and animation, while other testers found Opus 5.5 the best model ever tested for one-shot HTML generation.