Anthropic Ties Claude Opus 5.5 Pricing to Longer Coding Sessions

Anthropic reported on September 24, 2026 that Claude Code sessions have grown longer and more context-heavy over the past six months, detailing in a post on the Claude blog how Claude Opus 5.5’s pricing and caching mechanics address that shift. The company puts Opus 5.5’s running cost at roughly 40% below Opus 5 for a typical token-billed workload, its own estimate.

Six Months of Claude Code Usage Data

The post, written by Michael Segner, draws on aggregate usage data covering March through September 2026. The number of prompts per session held steady, the post reports, while the work inside each prompt grew: the model now works 3.3 times longer per prompt and makes more than 40% more model calls on each one, and interruptions fell 68%. Developers now connect a tool server or use a skill about twice as often, and paste text into prompts a third less often.

The amount of context in each request grew 2.6 times over the period, and the input-to-output token ratio moved from 189:1 to 324:1. Anthropic reads the numbers as developers aiming a longer-working, better-informed model at bigger, more open-ended tasks, and says the savings from Opus 5.5 are greatest for exactly those longer, higher-context sessions when usage is billed by token.

The analysis builds on Anthropic’s release of Claude Opus 5.5, which arrived with lower prices than Opus 5. Where the launch set the prices, the September 24 post and a companion analysis published September 22, 2026 by Addy Osmani examine what those prices mean for real Claude Code workloads.

Token Prices and the Cache-Read Cut

For usage billed by the token, Anthropic reduced input and output token prices 20% and cut the price of reading a cached token 60%. The company says the cache-read reduction carries the most weight because cache reads account for most of the cost of agentic and coding work, and it says that as of publication a cached token on Opus 5.5 comes in at one-fifth the cost of competing models and outperforms them.

Osmani’s companion analysis lists Opus 5.5’s API list prices at $4 per million input tokens, $20 per million output tokens, and $0.20 per million cache reads. At those prices, a cached read costs 5% of a fresh input token, while writing to the cache costs 1.25 times the input price for a five-minute cache and twice the input price for a one-hour cache, with each hit renewing the lifetime for free. On Pro, Max, or Team plans, the lower Opus 5.5 price is passed on to plan limits so they go about 25% further than on Opus 5; the additional cache-read discount applies only to API pricing.

Harness Changes That Protect the Cache

Even as context per request rose roughly 2.6 times, the share of input that misses the cache fell by more than 50% over the six-month window, the post reports. It credits a series of Claude Code changes that reduce accidental cache breaks, covering small disruptions such as a login refresh and larger ones such as adding instructions mid-conversation or loading tools on demand. On Opus 5.5 and Fable 5.1, developers can also change effort levels mid-session without resetting the cache.

The official prompt-caching documentation specifies that the effort-level behavior applies with an API key or a Claude subscription, and not on Amazon Bedrock, Google Cloud’s Agent Platform, or a Claude apps gateway, nor when the CLAUDECODEDISABLEEXPERIMENTALBETAS flag is set or an organization has a HIPAA configuration.

Developers on API keys and cloud providers can now set a one-hour cache lifetime, which subscribers already had. The documentation shows the defaults: one hour for the main conversation on a Claude subscription within plan usage, and five minutes on usage credits, API keys, or cloud providers. The promptCacheTtl setting or the CLAUDECODEPROMPTCACHETTL environment variable selects the longer lifetime, and both controls require Claude Code v2.1.242 or later.

Forked subagents now begin from the parent session’s cache rather than paying to reprocess the same context. The documentation explains that a fork takes over the parent’s system prompt, tool set, and full conversation history, so its opening request reads straight from the parent’s cache.

Fewer Turns Per Task

Anthropic reports that Opus 5.5 can complete a task in fewer turns than other models. According to the post, Zeta Labs recorded fewer turns and tool calls per task than on Opus 5, at nearly half the cost, and finished twice as many of its hardest tasks.

The post cautions that the pattern will not hold for every task, quoting Osmani’s analysis: “On a well-scoped task, both models finish in about the same number of turns, and the price cut is all you get. The gap should be biggest on open-ended tasks, where a model can spend many turns on the wrong idea.”

Anthropic also reports that Opus 5.5 produces output more than 30% faster than Opus 5. The gain leaves token use and cache hit rate unchanged; the post notes it shortens the wait on long runs as Claude spends more time working unattended.

The post closes with practices for protecting cached reads: run /usage in Claude Code to see what share of a session is being served from cache, choose the model when a session starts instead of switching partway through, run compaction before stepping away instead of afterward, and on an API key or cloud provider, enable the one-hour cache lifetime for long sessions.