What changed
Anthropic released Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 output, down 20% from Opus 5. Cache reads dropped from fifty cents to twenty cents per million tokens, a 60% cut. Anthropic says typical workloads now cost 40% less because the model uses fewer tokens and cache reads dominate long jobs. Opus 5.5 performs at Fable 5.1 level on most work while costing less than half.
What you can do with it now
Use Opus for long coding sessions, audits, migrations, and any job where you send the same large context repeatedly. Cache reads are the hidden cost in agent work: every time the model re-reads your system prompt, file tree, or reference docs, you pay cache rates, not fresh input rates. At twenty cents per million cached tokens, a 200,000-token codebase costs four cents per read instead of ten. Over a 30-turn coding session, that is $1.20 instead of $3. Set it as your default for any task that runs longer than five back-and-forth turns.
What to keep in your own hands
Don't let cost pick your model for the first pass. A model that takes three tries costs more than one that gets it right. Test Opus and Sol on the same task with actual work, not toy examples. Time the job, count retries, total the bill. Cache savings only matter if your tool actually preserves the cache between turns — check your setup before you trust the math. Never send your entire codebase to any model without stripping client names, API keys, and internal pricing first.