The number I care about is the cache reads, not the headline price. Most of my Claude use is agent work: the same long system prompt and repo context, fed in again and again while the model pokes at an Astro build or a Cloudflare config. That cached input just got 60 percent cheaper, and the model supposedly spends tokens more carefully and runs 30 percent faster.
What I’d do: rerun a week of my usual loops on Opus 5.5 and compare the bill. And watch the output. Anthropic says some requests it flags get quietly rerouted to an older, weaker model. That’s the kind of thing you want to catch before you blame your own code.
The story — Anthropic has released Claude Opus 5.5, claiming performance on par with Claude Fable 5.1 at roughly 40 percent less cost than Opus 5. Pricing is $4 per million input tokens and $20 per million output, with cache reads cut 60 percent. Subscribers get 20 percent higher five-hour limits. Automatic filters reroute suspect cybersecurity requests to an older, weaker model. Sonnet 5.5 and Haiku 5.5 follow in coming weeks (Source).