Claude Opus 5.5 brings pricing and usage changes, Anthropic says

How the API bill is calculated

Anthropic lists standard Claude Opus 5.5 API rates of $4 per million input tokens and $20 per million output tokens. Its September 22 announcement says typical workloads at default settings cost 40% less than Opus 5 in the company’s tests. That estimate combines token prices and the amount of work the model performs; it is not a guaranteed reduction for every application.

For a concrete example, assume a request uses 100,000 uncached input tokens and generates 10,000 output tokens. At standard rates, the token charge is $0.40 + $0.20 = $0.60. The same token counts at fast-mode rates would cost $0.80 + $0.40 = $1.20. These are illustrative calculations in U.S. dollars, not measured workloads. They exclude tool charges, taxes and other billing adjustments, and assume standard global routing without batch discounts.

Cache savings depend on reuse

The $5-per-million cache-write price applies to the five-minute cache. A one-hour write costs $8 per million tokens; cache reads cost $0.20 per million at standard speed. Writes and reads have different jobs: a write stores reusable prompt content, while a later matching request can read it at the lower rate.

Consider only a reusable 100,000-token prompt prefix across ten requests. One five-minute write costs $0.50, and nine successful cache reads cost $0.18, making that prefix’s total input charge $0.68. Sending it ten times without caching would cost $4.00. New input, generated output and tool use are additional in either case.

That example assumes every repeat request hits the same valid cache entry. Anthropic’s documentation requires an identical cached prefix and calls within the cache lifetime. Changed content, expired entries or different settings can require another write. The useful number to check in production is the recorded cache-hit usage, not the number of requests that were intended to reuse a prompt.

When fast mode is worth considering

Anthropic describes fast mode as the same model with a faster inference configuration. Its stated gain is up to 2.5 times as many output tokens per second, rather than a promise that an entire task finishes 2.5 times faster. The API feature remains a research preview with access conditions and separate rate limits. Switching between standard and fast mode also invalidates the prompt cache.

For a developer, that makes fast mode a latency decision: compare the extra charge with the value of receiving generated output sooner, while measuring time spent waiting for tools and reviewing results. For subscribers, the announced increases to five-hour allowances and the saveable rate-limit reset are separate from the API arithmetic above. Check the limits shown in the relevant account rather than converting API token prices into a promised number of messages.

Keep launch claims in their original context

The September 22 announcement also described a life-sciences verification route and a planned expansion of cybersecurity access. Those were launch-time statements; the pricing calculation does not establish today’s eligibility for specialized research. Anthropic described action screening, a sandbox and code review as safeguards, while acknowledging limits to its alignment testing. This article does not independently evaluate those safeguards.

Update history

Updated 13 October 2026 (UTC): expanded the original pricing report with five-minute and one-hour cache distinctions, reproducible cost examples and the documented limits of fast mode. The original publication date is unchanged. The announcement and current pricing, caching and fast-mode documentation were reviewed; no model benchmark or paid API workload was run for this update.