Claude Opus 5.5 pricing: calculate the whole agent run

Anthropic lowered the Claude API price of Opus 5.5 when it released the model on 22 September 2026. That matters to teams running coding agents, but a lower token rate does not tell you the cost of a completed change. An agent may read the same repository context many times, write to a cache, retry a tool call, or need a second run after review. Budget the job, not a single prompt.

What changed in the API price

As of 24 September 2026, Anthropic lists these standard prices per million tokens:

The input and output rates are 20% lower; the cache-read rate is 60% lower. Anthropic estimates that Opus 5.5 costs about 40% less on a typical workload at default settings. That 40% is its workload estimate, which also depends on token use. It is not a discount you can apply to every existing invoice. Fast mode has separate input and output rates of $8 and $40 per million tokens, so keep it out of a standard-speed comparison.

These are Claude API rates. A Claude subscription, a third-party coding product, or a negotiated enterprise agreement can have a different bill or usage limit. Check the route that actually runs your agent.

Turn a task log into a budget

For a Claude API run, multiply each billed token category, in millions, by its rate, then add the categories. For example, the Opus 5.5 standard-rate estimate is:

4 × input_millions + 20 × output_millions + 0.20 × cache_read_millions + 5 × five_minute_cache_write_millions + 8 × one_hour_cache_write_millions

That is only the model charge represented by those categories. Include any other billed services separately. Count a retry as part of the same attempted job; do not quietly remove failed runs when calculating cost per accepted change. If your agent tool reports only a total charge, use that observed charge rather than reconstructing it from a partial token log.

A useful pilot row records the starting task, model and effort, input/output/cache usage or billed charge, elapsed time, number of retries, reviewer minutes, repair minutes, and whether the change was accepted. Divide the total spend by accepted changes only after you have included abandoned attempts. Keep the task set and review standard the same when comparing Opus 5 with Opus 5.5.

Decide where the price reduction helps

Cache reads may be a large share of a coding agent's token traffic. A lower cache-read rate helps most when the agent actually reuses context and the provider bills those reads as cached tokens. A job that repeatedly generates large outputs, misses the cache, or needs substantial human repair can have a different result. Check the usage breakdown before predicting savings.

Anthropic also reports stronger coding and knowledge-work results for Opus 5.5, but its own release notes warn that small benchmark leads may not predict real-world differences. Set an acceptance bar on your repository: working tests, maintainable changes, safe recovery, and review time. Then compare the cost of the work that passes it.

If you are changing an existing agent, read the Opus 5.5 migration checklist before you switch its model ID. The coding-agent pilot scorecard covers a fair comparison with GPT-6 Sol.

Editorial note: This is a source-researched cost guide, not a measured benchmark or a claim about our own API spend. Prices and Anthropic's workload estimate were checked against Anthropic's Opus 5.5 announcement and Claude Platform pricing on 24 September 2026. Verify current rates and your billing terms before purchasing.