Claude Fast Mode on Opus 4.8 Ends November 30: What to Change, Plus 128 MB Requests and a New Usage API

Claude fast mode on Claude Opus 4.8 is deprecated, and Anthropic removes it on November 30, 2026. After that date, Opus 4.8 requests that ask for fast mode return an error. They do not quietly drop back to standard speed.

Anthropic announced the change in its October 30 platform release notes. The same week brought two smaller changes: a larger maximum request size and a new beta API for usage and cost reporting. This piece covers all three. The focus is what to change in code before Claude fast mode ends on Opus 4.8.

The short version: search your code for speed: "fast" on claude-opus-4-8. Move those calls to Claude Opus 5.5, which Anthropic names as the replacement. Or remove the fast setting and accept standard speed. Claude fast mode on Opus 5.5 costs less per token. But Opus 5.5 runs thinking on every request, so test before you switch. Separately, Messages requests can now be 128 MB on Anthropic’s own API, and a new Usage and Cost Analytics API can group spend by speed, which helps you find any fast mode traffic you missed.

What Anthropic changed for Claude fast mode

Anthropic’s platform release notes say fast mode for Claude Opus 4.8 was deprecated on October 30, 2026, "with removal on November 30, 2026." The entry continues: "After removal, requests to claude-opus-4-8 with speed: "fast" will return an error."

The fast mode documentation repeats the date and the error behavior. It describes fast mode as the same model running "with a faster inference configuration," with up to 2.5x higher output tokens per second at premium prices.

Two points are easy to miss:

• The model itself is not retiring. Anthropic’s model deprecations page lists claude-opus-4-8 as active, with retirement "Not sooner than May 28, 2027." Only the fast setting goes away.
• Nothing falls back for you. The docs say requesting speed: "fast" on a model that does not support it returns an error. Your code has to make the change.

If you need a refresher on the model, our Claude Opus 4.8 launch post covers what shipped in May, and our Opus 4.8 vs Opus 5 comparison covers the step after it.

Find every Claude fast mode call before November 30

Claude fast mode was a research preview on Opus 4.8, so it may sit in only a few places. Search anyway. A chatbot that fails on December 1 is harder to trace.

Look for these in code, config files and environment variables:

• speed="fast" or "speed": "fast" in request parameters.
• The beta header fast-mode-2026-02-01. The fast mode docs list it as required for Opus 4.8 and Opus 5, and not for Opus 5.5.
• Plugin or internal tool settings that pass a "fast" toggle to the API.

Then check traffic. Code search misses vendor tools you did not write. The fast mode docs say each response’s usage object includes a speed field, set to "fast" or "standard". If you log usage, filter for "fast" with model claude-opus-4-8. The new analytics API, covered below, can do the same across a whole organization.

What to use instead, according to Anthropic

The release notes give one path: "To continue using fast mode, migrate to Claude Opus 5.5." The fast mode page lists Claude fast mode as available on Opus 5.5. On Opus 5 and Opus 4.8 it is a research preview. Opus 5 access needs an account manager or a waitlist, per the same page.

Prices from Anthropic’s pricing page, as of early November 2026:

Per token, Claude fast mode on Opus 5.5 is 20% cheaper than on Opus 4.8. The per-request cost can still go up. Anthropic’s Opus 5.5 migration guide says that on Opus 4.8, thinking is off unless you ask for it. On Opus 5.5, a request with no thinking field runs with thinking, and thinking tokens bill as output.

You have two other choices. You can keep Opus 4.8 at standard speed by removing the speed parameter and the beta header. Or you can test whether a faster, cheaper model does the job at standard speed. Our Opus 5.5 launch post covers how its effort settings change cost.

Moving a fast mode call to Opus 5.5

The model swap is the small part. The migration guide lists settings that Opus 5.5 rejects with a 400 error. Opus 4.8 code often uses some of them:

• thinking: {"type": "disabled"} and manual thinking budgets.
• Forced tool choice ("any" or a named tool).
• temperature, top_p and top_k.
• A prefilled assistant turn at the end of the messages.
• The older computer_20251124 tool on the Claude API.

The guide also notes that Opus 5.5 defaults to medium effort, while Opus 4.8 defaults to high. Opus 5.5 does not support Priority Tier. The fast mode page adds that Claude fast mode does not work with a Priority Tier commitment or the Batch API.

We walked through each of these breaks in our Claude Opus 5.5 migration guide. If you are also moving SDK versions at the same time, read our post on the new Claude API SDK major versions first and change one thing at a time.

Rate limits and fallback for Claude fast mode on Opus 5.5

According to the fast mode docs, Opus 5.5 has a default fast mode rate limit in the Start, Build and Scale usage tiers, so no access request is needed. Organizations in the Evaluation tier get it after moving up a tier.

Fast mode has its own rate limit. Going over it returns a 429 with a retry-after header. The docs show a fallback: catch the error and retry without speed: "fast".

That fallback has a cost the docs call out. "Falling back from fast to standard speed results in a prompt cache miss." Requests at different speeds do not share cached prefixes. If your prompts rely on a large cached system prompt, a fallback request pays full input price.

Bigger requests: 128 MB on the Claude API

On October 28, Anthropic raised the maximum request size for the Messages API and the Token Counting API from 32 MB to 128 MB. The release notes limit this to "the first-party Claude API."

The API overview lists the current limits: 128 MB for Messages and Token Counting, 256 MB for Message Batches, 500 MB for the Files API, and 32 MB for Sessions, Agents and Environments. Going over returns a 413 request_too_large error.

In practice, this allows larger inline PDFs, more images per request, and long document sets sent as base64. Base64 makes a file about a third larger. The new limit leaves room for scanned board packets or image-heavy reports.

Partner platforms did not change. The same page says Bedrock limits requests to 20 MB and Claude Models on Google Cloud (Google managed) to 30 MB. Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry stay at 32 MB. If one codebase serves several platforms, check size against the smallest limit you deploy to.

Switch reporting to the Usage and Cost Analytics API

On October 29, Anthropic launched the Usage and Cost Analytics API in beta. It has two endpoints, /v1/organizations/analytics/usage and /v1/organizations/analytics/cost. Requests need the beta header analytics-usage-cost-2026-08-24.

The same day, the older Usage and Cost API became a legacy API. So did four report endpoints of the Claude Enterprise Analytics API. The release notes say "All of them remain supported," and the docs give no shutoff date as of early November 2026.

The docs list what changes when you move a Console integration:

• Freshness: legacy data "typically appears within 5 minutes." The new API "typically refreshes within 4 hours and can take up to 24 hours."
• Shape: rows ranked by order_by, with no rows for empty buckets. bucket_width becomes time_bucket_width, with no default.
• Cost units: amounts are strings in cents with six decimals. The docs give the example that "41280.000000" is $412.80.
• Range: a request can span at most 31 days and look back 365 days.
• New groupings: actor, context_window, service_tier, inference_geo and speed.

That last one ties back to the deadline. Group cost by speed and filter models[]=claude-opus-4-8 to see whether any Claude fast mode spend remains. Data can lag a day, so run the check again in the last week of November.

Keep the legacy integration running while you test the new one. Budget alerts built on five-minute data will be slower. Our post on Claude OpenTelemetry tracing covers per-request tracing, which fills that gap.

How OpenAI and Google handle the same jobs

OpenAI’s API has its own Costs endpoint under Administration. It groups by project, user, line item, API key and API source, in daily buckets only. Our OpenAI model shutoff post covers its October shutdown of seventeen models. Google took a different approach for some models. Its Gemini API changelog says requests to the deprecated Gemini 3.7 Flash and Gemini 3.5 Flash are routed automatically to newer models. Our post on Gemini shutdown dates lists the Google models that do switch off.

Frequently Asked Questions