Claude Opus 5.5 pricing: what the 40% cost reduction means

Claude Opus 5.5 costs $4 per million ordinary input tokens and $20 per million output tokens on Anthropic’s direct API. Those rates are 20% below Opus 5. The larger 40% figure in the launch announcement describes Anthropic’s estimate of typical task costs, combining lower rates with changes in token use.

This distinction matters when you budget a coding agent or compare an old invoice with a new run. The calculations here use the Anthropic announcement and pricing documentation, checked September 23, 2026. Prices are USD per million tokens, excluding separate applicable charges. They are not an Ofox quote.

Separate the three reductions

A workload with substantial cache reads has a different effective reduction from one that consists mostly of new input and output. A model that uses fewer tokens can reduce a task’s cost further, but your workflow may not reproduce the manufacturer’s test mix.

Keep the claim attached to its source: Anthropic reports an estimated 40% reduction on typical workloads at default settings. This is not a guaranteed discount on every bill, every subscription or every provider route. It also does not mean that all five fields in the table were cut by 40%.

Recalculate one fixed workload

Assume a request is billed for 20,000 ordinary input tokens, 60,000 cache-read tokens, 20,000 five-minute cache-write tokens and 5,000 output tokens. These categories do not overlap. The example deliberately holds token counts fixed to isolate the rate change.

The fixed workload costs about 23.2% less at the new rates. There is no contradiction with a manufacturer’s separate 40% task-cost estimate: this example does not assume a reduction in tokens consumed. It is an accounting illustration, not an Opus performance test.

Download the rate sheet and calculator to reproduce the numbers. Enter observed usage when you evaluate your own application. Do not substitute the number of visible words for billed output tokens.

Cache duration changes the write price

The five-minute and one-hour cache-write rates are different. Choose the duration that matches the pattern of reuse, then include its initial write cost. Repeated reads can justify a longer-lived cache, but an input prefix used only once cannot deliver repeated-read savings.

As with the cache accounting worksheet, ordinary input, reads and writes must be mutually exclusive in the cost formula. If an export reports total input plus cache details, interpret those fields before multiplying anything. Counting a cached prefix in both total ordinary input and cache-write input inflates the bill estimate.

Do not transfer OpenAI’s 30-minute caching rules or long-context threshold into an Anthropic request. Similar billing labels do not establish identical product rules.

Fast and Batch answer different needs

Opus 5.5 Fast mode lists $8 input and $40 output per million tokens. Anthropic describes a faster service option; that does not make every complete agent task faster by a fixed amount. Tool execution, repository size and review work also take time.

Batch lists $2 input and $10 output for asynchronous work. Fast and Batch are not a combined mode. Cache duration, supported processing options and regional requirements must be checked against the current rate documentation before estimating a production workload. A table of two token rates alone is not enough to price every combination.

For a user waiting on an interactive task, compare the value of lower latency with the higher rate. For an offline job, assess whether asynchronous processing satisfies the deadline. Keep these decisions separate from the model’s ability to produce an acceptable answer.

Subscription limits are another system

The launch also discusses increased five-hour usage limits and a rate-limit reset that subscribers can save. Those statements do not define an unlimited plan or a fixed conversion from subscription messages to API tokens.

The Claude Code access guide explains model selection and usage checks. The Sol versus Opus 5.5 guide compares API task costs with explicit assumptions. Use an actual provider quote if you intend to buy through a gateway; direct Anthropic prices do not certify that gateway’s support or rate.