Anthropic published a detailed cost breakdown of its Opus 5.5 model, arguing that the number that matters to a developer is not the price per token but the price per finished task. The post, written by Addy Osmani, walks through how turns, cache reads, and output tokens combine inside Claude Code, the company’s agentic coding tool, to produce a final bill that can vary widely between two sessions using the identical model.
The headline numbers are straightforward, if you know where to look. Cache reads list at $0.20 per million tokens, output at $20 per million, and input at $4 per million, list prices Anthropic itself published for Opus 5.5. Against Opus 5, that is a 20 percent cut on input and output and a steeper 60 percent cut on cache reads. Anthropic says that second discount compounds inside Claude Code, since most of what a tool-calling session resends on each turn is text the model already processed once and is now pulling from cache.
That distinction between “cheaper tokens” and “cheaper tasks” is the real argument of the post. Anthropic’s own worked example shows a task that runs 40 turns over a conversation that grows from 20,000 to 120,000 tokens costs roughly $1.62 in input alone at a 90 percent cache hit rate. The same task finished in 25 turns costs about $1.02, even though the model, the price list, and the final conversation length never changed. The company’s framing: a turn resends everything that came before it, so the turn a developer avoids is worth more than any discount on the tokens inside it.
Output tokens carry the sharpest multiplier in the whole system. Anthropic says an output token on Opus 5.5 costs 100 times a cache read, and that thinking tokens (the model’s internal reasoning before it answers) are billed as output even when Claude Code shows the user only a summary. That is why the company ties “effort,” the setting that controls how much a model reasons per turn, directly to spend: raising effort from medium to high can add tens of thousands of thinking tokens to a single task.
Anthropic frames the tradeoff for operators bluntly: every lever that saves tokens (lower effort, a smaller model, less context) can also cost a finished task, and a single retry loop can outspend all of those savings combined. The post’s own arithmetic makes the case. It estimates that raising effort by roughly 20,000 thinking tokens costs about $0.40 on Opus 5.5, which is close to what a 10-turn retry loop on 100,000 tokens of cached context would cost. In other words, higher effort pays for itself the moment it prevents even one retry, and is wasted the moment it doesn’t.
The post is Anthropic’s own analysis of its own product, run on its own benchmark: a 44-ticket internal customer support test. Moving that workload onto the newer model, set to low effort, brought the bill down by roughly 18 percent versus the older Opus 4.8 baseline. Layering on the company’s own prompt-audit tool trimmed a further 9 percent, landing close to 25 percent below where Opus 4.8 started. Anthropic has not published comparable cost breakdowns for competing models on the same benchmark, so the savings figures describe Anthropic’s product against its own prior version rather than against rival coding agents from OpenAI or Google.
Anthropic’s closing advice reads more like an internal runbook than a pitch. Well-scoped, routine work should stay at medium effort; only escalate once a test, a build step, or a live endpoint call shows the model actually needs to think harder. Lookups, log reads, and other search-heavy subagent work belong on a cheaper model such as Sonnet or Haiku, not on the model that is writing the code. Whatever the setup, Anthropic points developers to the /usage command at the end of a session, the only way to see the real split between cache reads, fresh input, and output instead of guessing.
For any team already metering Claude Code spend against a budget, the actionable move is not to chase the lower per-token price but to run the same real task twice, once at each effort level, and compare the /usage output before deciding whether Opus 5.5’s cheaper rate card changes anything for that specific workflow.
Based on reporting and figures published by Anthropic on the Claude blog, September 2026.
