An Anthropic blog post by Michael Segner, dated 2026-09-24, about how Claude Code sessions changed between March and September 2026 and how Claude Opus 5.5 is priced and tuned for them.

Caveat: this is vendor marketing. Every number is Anthropic-reported, with no method, sample, or workload definition given; “typical workloads” is not defined. The post does not list absolute prices.1

Headline numbers

ClaimFigure
Cost to run versus Opus 5, typical token-billed workloadsabout 40% less
Input and output token price20% lower
Cached token read price60% lower
Input missing the cachedown more than 50%
Output speed versus Opus 5over 30% faster

As reported in the post.1

Takeaways

  • Sessions got longer and heavier on context, so cached reads now dominate agentic cost. See Claude Code session cost.
  • Savings come from lower prices, fewer cache misses, and fewer turns. See Where the savings come from.
  • Subagents now start from the parent’s cache. See Subagents.
  • Watch cached reads with /usage, pick the model at session start, and use a one-hour cache for long API sessions.1 See Keeping the cache warm.

Footnotes

  1. Claude Opus 5.5 and longer coding sessions, original ↩ ↩2 ↩3