# Claude Opus 5.5 Fast Mode: API Pricing, Access and When to Use It

Anthropic has removed the fast-mode waitlist, but deciding which requests deserve the premium takes more than comparing token speeds.

By Paras Madan | 2026-10-17


Anthropic moved Claude Opus 5.5 fast mode out of research preview on October 15, 2026. Developers can now use the faster inference option without requesting access. It gives applications a way to reduce generation time while keeping the same model, at a higher token price.

The practical change in the [release notes](https://platform.claude.com/docs/en/release-notes/overview) is access: Start, Build and Scale tiers receive default fast-mode limits, and Opus 5.5 no longer needs the preview beta header. Evaluation-tier organizations need to move to an eligible tier. Older supported Opus models remain in preview.

## Claude Opus 5.5 fast-mode pricing

Anthropic's [pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing) lists these US-dollar rates per million tokens:

| Opus 5.5 option | Input | Output |
| --- | --- | --- |
| Standard API | $4 | $20 |
| Fast mode | $8 | $40 |
| Asynchronous Batch API | $2 | $10 |

Fast mode costs twice the standard base rate. Caching and data-residency modifiers still apply, and fast mode cannot be combined with the Batch API. The fast rates cover the full context window, including requests above 200,000 input tokens.

For a simple illustration, 10,000 uncached input tokens and 2,000 output tokens cost $0.08 at standard speed or $0.16 in fast mode, before other charges. That difference looks small for one response. Across an agent that repeatedly reads context and generates tool calls, it becomes a routing decision worth measuring.

## What gets faster

Anthropic describes up to 2.5 times higher output tokens per second. Its [fast-mode guide](https://platform.claude.com/docs/en/build-with-claude/fast-mode) says the weights and capabilities stay the same; the inference configuration changes. The improvement targets generation throughput, rather than time to first token.

This makes the shape of the workload important. A long answer streamed to someone waiting at a screen is a plausible candidate. A short classification request may have little generation time to save. An agent waiting on a database or a slow external tool may remain dominated by that dependency.

**Measure the time to finish the user's task.** Token throughput is only one part of that measurement, and the advertised maximum does not establish the gain your application will see.

## How to enable fast mode in the API

Use model ID `claude-opus-5-5` and add `speed: "fast"` to the Messages request. Check `usage.speed` in the response to record the speed used.

Fast mode has separate rate limits. A `429` can mean fast capacity is unavailable even when the standard route has capacity. Decide whether that request should wait or fall back, rather than leaving the user-facing behavior accidental.

There is also a cache tradeoff: fast and standard requests do not share cached prefixes. Switching speeds invalidates the prompt cache. For a conversation with a large repeated context, alternating routes can undermine the savings expected from caching.

A useful implementation test is to compare three complete sessions: standard throughout, fast throughout, and selective fast routing. Track elapsed time, retries, cache reads and the final bill. That gives the team a better basis for choosing than a single fast response.

## Availability and Claude Code

The [official announcement](https://x.com/ClaudeDevs/status/2110825294655562048) also confirms availability in Microsoft Foundry deployments hosted on Azure. Google Cloud is described as coming soon. Anthropic-hosted Foundry deployments and Claude Platform on AWS are not supported by the current fast-mode guide.

Claude Code's fast mode remains unchanged and uses extra usage credits. Removing the API waitlist does not turn it into an included subscription benefit.

For background work with no person waiting, the Batch API's lower price remains a useful comparison. For interactive work, the question is whether less waiting produces enough value to justify both the premium and the operational complexity. The release makes that choice easier to access. The application's own measurements should decide where it becomes the default.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and product updates into clear stories developers and founders want to read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min)