Claude Opus 5.5 fast mode: pricing and how to turn it on

Claude Opus 5.5 fast mode costs $8 per million input tokens and $40 per million output tokens, twice the standard Opus 5.5 rate, and it writes output up to 2.5 times faster. As of October 15 there's no waitlist and no beta header: you add speed: "fast" to an API request and that's it.

Anthropic's developer account made the announcement. @ClaudeDevs said Opus 5.5 fast mode is "no longer in research preview, and we've removed the waitlist," and described it as the same model running faster. The post quoted the account's own May 12 post, which opened fast mode as a research preview for Claude Opus 4.7. So it took about five months for fast mode to drop the access gate, and it's Opus 5.5 that got there first.

What Opus 5.5 fast mode costs

Anthropic's pricing page lists standard Opus 5.5 at $4 in and $20 out per million tokens. Fast mode doubles both. Opus 5 and Opus 4.8 also have fast mode, at $10 in and $50 out, but they're still in research preview and you have to request access.

Some details that change the bill:

• The fast price covers the whole context window, including requests over 200,000 input tokens.
• Prompt caching discounts and data residency multipliers apply on top of the fast rate.
• Fast mode doesn't work with the Batch API, so it can't be combined with batch's half-price rates.
• It isn't available with a Priority Tier commitment or on Claude Platform on AWS.

On a single reply, the gap is small. A 2,000-token answer costs 4 cents at standard speed and 8 cents in fast mode. Multiply that by an agent making thousands of calls a day and it starts to show up.

How to turn it on

According to Anthropic's fast mode docs, you set speed to "fast" on a Messages API call with model claude-opus-5-5. Every official SDK supports the parameter. Managed Agents take the same setting as model.speed. If your code still sends the old fast-mode-2026-02-01 header, it keeps working.

To confirm which speed you actually got, check the speed field in the response's usage object. It reads "fast" or "standard".

Access depends on your usage tier. Opus 5.5 has a default fast mode rate limit in the Start, Build and Scale tiers. Organizations still in the Evaluation tier have to move up a tier before they can use it. Fast mode has its own rate limit, separate from your normal Opus limit, and the Console's rate limits page shows how much of it you have. Go over it and you get a 429 error with a retry-after header.

Besides the Claude API, it runs on Claude in Microsoft Foundry for deployments hosted on Azure. The ClaudeDevs post said Google Cloud support is "coming soon."

The catches worth knowing

The speedup is in output tokens per second. Time to first token doesn't change. A long streamed answer, like a big code edit or a full report, finishes much sooner. A short classification that returns ten words barely changes, and you still pay double.

There's also a caching problem. The docs say falling back from fast to standard speed causes a prompt cache miss, because the two speeds don't share cached prefixes. Anthropic's sample code catches the rate limit error and resends the request without speed: "fast". That keeps the job running, but you pay full input price to rebuild the cache. If your agent sends a large cached system prompt on every call, switching speeds back and forth could cost more than the speedup saves.

Claude Code hasn't changed. @ClaudeDevs said fast mode there "remains unchanged, and uses extra usage credits." So the waitlist is gone for API customers, but subscribers who use fast mode in Claude Code still pay for it as extra usage beyond their plan.

Our read

Anthropic says fast mode runs the same weights with a faster inference setup, so you aren't trading quality for speed. You're paying for time, and the decision comes down to how much a minute of waiting costs you. For someone watching a cursor in a chat window, or a sales agent drafting a reply while the customer is still on the phone, double the price may be easy to justify. Overnight jobs that nobody watches should stay on standard speed or go through batch.

What this means for a business using AI

Most automations in a small business run in the background: sorting leads, drafting follow-ups, summarizing calls. Paying double for speed there doesn't buy you anything. The question is worth asking only where a person is waiting on the answer, and even then, try it on a week of real traffic before you commit.

If you're not sure which of your workflows need a fast model and which can run on the cheap, slow setting, New Face Design's free process audit maps where your hours go and shows where automation will pay for itself.