A claude.dev blog post by Addy Osmani, published Sep 25, 2026, snapshotted the same day. It explains what sets the cost of a Claude Code task, what the Opus 5.5 price cut changes, and which settings (effort, model, caching, compaction) trade tokens against a finished task.
Scope and gaps: most dollar figures are illustrations the author built from API list prices, and the post says so. The interactive calculators (Fig A to C) were captured only as the default values shown in the page text. The “40% less” headline is Anthropic’s estimate for typical workloads, which the post separates from the per-token price change. The prompt-audit result comes from one internal benchmark of 44 tickets.1
Takeaways
- A task is a loop, and four things set its cost: turns, cache reads, output tokens (thinking included), and the model. See Claude Code session cost.
- Opus 5.5 cuts input and output prices by 20% and cache reads by 60% against Opus 5; the illustrative session in the post costs about 31% less from price alone. See Opus 5.5 pricing.
- Raise effort before changing models; Opus 5.5 defaults to medium. See effort.
- Use Sonnet or Haiku subagents for lookups, Opus 5.5 for supervised work, Fable 5.1 for the hardest unsupervised runs. See model choice and Subagents.
- Keep the cache warm, compact at a break, and measure with
/usage. See caching and compaction.1