Claude Opus 5.5 Cuts Coding Costs 40% With Smarter Caching
Claude CodeAnthropic used the Claude Opus 5.5 launch to publish six months of usage data showing Claude Code sessions running 3.3 times longer per prompt with 2.6 times more context, and re-priced the model around that shift: input and output tokens are 20% cheaper and cache reads, which Anthropic says dominate agentic coding costs, are 60% cheaper, driving roughly a 40% total cost reduction on typical workloads. Anthropic also reworked parts of the Claude Code harness so effort-level changes and forked subagents no longer force a full cache rebuild, which it credits with cutting cache misses by more than 50%.
Featured Video
A video we selected to help illustrate this changelog
Key Takeaways
- Opus 5.5 cuts input and output token prices 20%, to $4 per million input tokens and $20 per million output tokens versus Opus 5.
- A typical coding workload costs roughly 40% less to run on Opus 5.5 than on Opus 5, while generating output about 30% faster.
- Cache reads, not raw generation, dominate the cost of agentic coding work, which is why Anthropic targeted a 60% cut there specifically rather than only cutting headline token prices.
- Anthropic's own usage data shows the input-to-output token ratio shifting from 189:1 to 324:1 over six months, a sign that sessions increasingly reread large context to make small edits.
- Effort-level changes mid-session no longer force a cache rebuild on Opus 5.5, closing a common source of wasted cache spend in long Claude Code sessions.
- Forked subagents now inherit their parent's cache instead of starting cold, directly cutting cost for workflows that spin up multiple subagents from one session.
Sources & Mentions
5 external resources covering this update
Opus 5.5 Live Test (Claude Code, Agentic Coding, Frontend Design, Brownfield Projects)
YouTube
Anthropic Launches Claude Opus 5.5 With Lower Prices and Faster Output
TechRepublic
Claude Opus 5.5 offers faster coding and improved AI performance
Android Authority
Claude Opus 5.5: Anthropic cuts price 20%, tests safety
BetaNews
Claude Opus 5.5 Review: Faster AI and Cheaper Pricing
Geeky Gadgets
Coding Sessions Have Changed, and Pricing Follows
Anthropic used the Claude Opus 5.5 launch to publish six months of data on how Claude Code sessions have actually evolved, and re-priced the model around what it found. Between March and September 2026, developers using Claude Code kept sessions running 3.3 times longer per prompt, with 40% more individual model calls inside a single request. The amount of context carried into each request grew 2.6 times over the same period, and the ratio of input tokens to output tokens shifted from 189:1 to 324:1, evidence that agentic coding work increasingly means rereading large amounts of context to produce comparatively small, targeted edits.
Cheaper Input, Much Cheaper Cache Reads
Opus 5.5 lowers input and output token pricing by 20% versus Opus 5, to $4 per million input tokens and $20 per million output tokens. Cache reads, which Anthropic says make up the bulk of the cost in agentic and coding work, dropped 60%, to $0.20 per million tokens down from $0.50. Combined with model-level efficiency gains, Anthropic says a typical workload on Opus 5.5 costs roughly 40% less to run than the same workload on Opus 5, while generating output about 30% faster.
Harder to Accidentally Break the Cache
Anthropic also rewired parts of the Claude Code harness so a long session is less likely to pay for a full cache rebuild by accident. On Opus 5.5 and Fable 5.1, changing the effort level mid-session no longer resets the prompt cache. Forked subagents now start from their parent session's existing cache instead of rebuilding one from scratch. Anthropic also says a range of smaller cache-breaking triggers, such as a login token silently refreshing mid-session, mid-conversation instruction changes, or tools loading on demand, have been tightened up so they no longer force an unnecessary cache miss. Anthropic reports these changes contributed to cache misses falling by more than 50% even as sessions and their context windows have grown substantially larger.