Cache hits cut the bill from $11.20 to $1.62 on the same Opus 5.5 task
When Anthropic launched Opus 5.5, it also published a cost exercise spelling out what a complete coding task actually costs — and the conclusion is that the sticker price is the smaller half of the story.
Cache hits cut the bill from $11.20 to $1.62 on the same Opus 5.5 task
When Anthropic launched Opus 5.5, it also published a cost exercise spelling out what a complete coding task actually costs — and the conclusion is that the sticker price is the smaller half of the story.
When Anthropic launched Opus 5.5, it also published a cost exercise spelling out what a complete coding task actually costs — and the conclusion is that the sticker price is the smaller half of the story. What determines the bill is the number of turns, the share of tokens read from the cache, and the shape of the workload.
The news: the arithmetic behind the price
Claude Opus 5.5 launched on September 22, 2026 with an input price of $4 per million tokens and $20 for output. Cache reads — tokens fetched from the prompt cache rather than reprocessed — received the largest cut, 60 percent, down to 20 cents per million, SiliconANGLE reported from the launch. Anthropic said at the same time that, according to the company, this amounts to 40 percent less on a typical workload than Opus 5 — an estimate that has not been independently verified.
The same day, Anthropic published the blog post "What a task costs on Opus 5.5", written by Addy Osmani. It states: "All prices are lower. Input and output tokens are 20% cheaper than on Opus 5. Cache reads are 60% cheaper."
The post's point is that the extra savings on cache reads apply to API pricing only, while subscribers on the Pro, Max, and Team plans get roughly 25 percent longer usage limits, according to the company.
The arithmetic: why turns dominate
The core of Osmani's walkthrough is a property of all chat-based models: every turn in a session resends the entire conversation to the model. As the conversation grows, each individual turn therefore sends ever more tokens — even if the context never becomes large.
The post works through a hypothetical example, which Anthropic explicitly labels a best-effort estimate and not a measurement: a task starts with 20,000 tokens of context and grows to 120,000 as the model reads files and tool results. Over 40 turns, each turn sends an average of around 70,000 tokens. That gives roughly 2.8 million input tokens for the task as a whole. With a 90 percent cache hit rate, the input costs about $1.62. The same task in 25 turns processes around 1.75 million tokens and costs about $1.02 in input.
The cache hit rate is the single factor that matters most. The same 2.8 million input tokens cost $11.20 if nothing comes from the cache, $1.62 at 90 percent hits, and around $0.99 at 96 percent. "No other setting moves the input price as much," Osmani writes.
The shape of the workload determines the savings
The savings against Opus 5 are therefore not fixed. According to the post, a session consisting mainly of cache reads can save up to 60 percent on input. A short question with no cache and a long answer, by contrast, can save only up to 20 percent, because output tokens — which get no benefit from the cache — dominate the bill.
Osmani also points to an interesting pattern in who benefits most: on a well-scoped task, both models use roughly the same number of turns, and then the price cut is all you get. "The gap should be largest on open-ended tasks, where a model can spend many turns on the wrong idea," the analysis states, as reproduced by Unite.AI.
The post also concedes a downside: every way of using fewer tokens can also cost a completed task. A lower effort level, a smaller model, or less context saves tokens — but a retry costs more than those savings.
The usage data behind the pricing logic
On September 24, Anthropic followed up with a post by Michael Segner on what Claude Code sessions actually look like. The figures are known via Unite.AI's account, so they should be read as the company's own aggregate usage data.
In the six-month window from March to September 2026, the number of questions per session held steady, while the work inside each question grew: the model now works 3.3 times longer per question, makes over 40 percent more model calls per question, and interruptions fell by 68 percent. Context per request increased roughly 2.6 times, and the ratio of input to output tokens moved from 189:1 to 324:1.
The most important point for the bill: even though context per request grew 2.6 times, the share of input that misses the cache fell by more than 50 percent over the period. Anthropic attributes this to a series of Claude Code changes that reduce unintended cache breaks — everything from small disruptions such as a login refresh, to larger ones such as adding instructions mid-conversation or loading tools on demand.
In other words: the model sends ever more data, but an ever larger share of that data is read from the cache at 20 cents per million instead of $4. This is what makes the cache hit rate the decisive cost variable — and it explains why Anthropic chose to cut precisely that price line the hardest.
The levers you can use yourself
For Claude Code users, there are concrete controls. The documentation states that the promptCacheTtl setting or the CLAUDE_CODE_PROMPT_CACHE_TTL environment variable selects a longer cache lifetime (one hour instead of the default), and that both require Claude Code v2.1.242 or newer. Mid-session changes to the effort level without a cache reset work with an API key or a subscription, but not on Amazon Bedrock, Google Cloud's Agent Platform, or a Claude apps gateway — and not with the experimental betas flag or in HIPAA configurations.
Another change: forked subagents now start from the parent session's cache, instead of paying to reprocess the same context. Since a subagent otherwise had to re-process the parent's system prompt, tool set, and conversation history, this saves exactly the kind of duplicated input that the arithmetic above shows gets expensive.
Anthropic also cites a customer result: Zeta Labs reportedly recorded fewer turns and tool calls per task than on Opus 5, at close to half the cost, and completed twice as many of its hardest tasks. This is a single-customer result relayed by Anthropic itself, without independent confirmation.
What remains
The picture is never as clean as the list prices suggest. "40 percent cheaper on a typical workload" is Anthropic's own estimate, not an independent measurement. The task costs in the blog post are illustrative arithmetic built on list prices — not benchmark data. The September 24 usage statistics rest on secondary sources via Unite.AI.
One unresolved point: SiliconANGLE mentions at launch a faster serving mode at $8/$40 per million tokens, which neither the blog post nor Unite.AI addresses — it should be treated as unconfirmed.
The firm impression that remains, nonetheless: for those using Opus 5.5 on coding work, the price per token is not the decisive variable. The number of turns a task requires, and how well the session retains its cache, move the bill more than anything else — and there are controls for both.