Claude Opus 4.8 fast mode ends November 30: where to move it

Anthropic deprecated fast mode for Claude Opus 4.8 on October 30, 2026, and will remove it on November 30, 2026. After removal, a request to claude-opus-4-8 with speed: "fast" returns an error. Teams using fast mode on Opus 4.8 can move that traffic to Claude Opus 5.5, the path Anthropic names, or drop back to standard speed before the cutoff.

The move looks cheaper on paper. Opus 5.5 fast mode lists at $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Opus 4.8. The catch is that Opus 5.5 thinks on every request and bills that thinking as output tokens, so a workload that ran without thinking on 4.8 can cost more per request even at the lower rate.

What changed in Claude Opus 4.8 fast mode?

Fast mode on Opus 4.8 now has an end date. The Claude Platform release notes for October 30 say fast mode for Claude Opus 4.8 is deprecated, with removal on November 30, 2026, and that teams who want to keep fast mode should migrate to Claude Opus 5.5.

Until removal, the fast mode documentation lists Opus 4.8 fast mode as a research preview on the Claude API that uses the fast-mode-2026-02-01 beta header. The notice covers fast mode only. Standard-speed Opus 4.8 still appears on Anthropic's pricing page at $5 and $25 per million input and output tokens.

How Claude fast mode works

Fast mode runs the same model on a faster inference configuration. Anthropic says it delivers up to 2.5x higher output tokens per second, with no change to intelligence or capabilities. The speedup is focused on output tokens per second, not time to first token, and the gain is most visible when streaming.

A few mechanics carry over to any migration:

• You opt in per request by setting speed: "fast". The response's usage.speed field reports whether the request actually ran at "fast" or "standard".
• Fast mode has its own rate limit, separate from standard Opus limits. When it runs out, the API returns a 429 with a retry-after header.
• Fast and standard requests don't share prompt cache. Anthropic notes that falling back from fast to standard speed causes a cache miss.
• Fast mode isn't available through the Batch API or with a Priority Tier commitment.

Claude Opus 5.5 fast mode pricing compared

Opus 5.5 is the only model where Anthropic lists fast mode as available rather than in research preview, and its fast rate is 20% lower than the Opus 4.8 fast rate. Prompt caching and data residency multipliers apply on top of fast mode pricing for both.

Scroll sideways to see all 3 columns.

Option	Input / output per million tokens	Notes
Opus 4.8, fast mode	$10 / $50	Research preview; removed November 30, 2026
Opus 5.5, fast mode	$8 / $40	Available; no beta header; adaptive thinking on every request
Opus 4.8, standard speed	$5 / $25	Not affected by this notice
Opus 5.5, standard speed	$4 / $20	Adaptive thinking on every request

Opus 5.5 fast mode also reaches more platforms. The pricing page lists it on the Claude API, on Claude Models on Google Cloud (Anthropic managed) and on Claude in Microsoft Foundry for deployments hosted on Azure. Opus 4.8 fast mode runs on the Claude API only. Opus 5.5 also gets a default fast mode rate limit in the Start, Build and Scale usage tiers, so you don't need to request access.

Does Opus 5.5 fast mode cost less in practice?

It costs less per token, and it may not cost less per request. The Opus 5.5 migration guide says that on Opus 4.8, requests without a thinking field run without thinking, while on Opus 5.5 every request runs with adaptive thinking, and thinking can't be disabled. Thinking tokens are billed as output tokens even when the thinking text isn't returned.

BaristaLabs' arithmetic: Opus 5.5 fast output costs $40 per million tokens against $50, so on output alone it breaks even when output tokens per request grow by 25%. Cheaper input moves that line higher, at $8 against $10. With equal input and output tokens, the break-even is 30% output growth. With 10 input tokens for every output token, it is 75%. These figures assume the same input tokens on both models and no prompt caching. A workload that returns short answers with no thinking today could still cross that line. A workload that already ran with thinking on 4.8 has less new thinking to absorb, so it may be more likely to come out ahead.

The guide names one request parameter for thinking depth: effort. Opus 5.5 supports low, medium, high, xhigh and max, with medium as the default, and a lower level spends fewer tokens on thinking.

Which requests need code changes before November 30?

Opus 4.8 fast-mode code can need more than a new model ID. The migration guide lists what every request to claude-opus-5-5 must satisfy. The thinking, tool choice, sampling and prefill settings below return a 400 error on Opus 5.5. Response parsing, output limits and the beta header don't:

• Thinking settings. thinking: {"type": "disabled"} and manual budget_tokens are rejected. Send no thinking field, or {"type": "adaptive"}, and use effort instead.
• Forced tool calls. tool_choice of {"type": "any"} or a named tool is rejected. Use auto, the default, or none.
• Sampling and prefill. Omit temperature, top_p and top_k, and don't end messages with a prefilled assistant turn, which is rejected.
• Response parsing. A response can start with thinking blocks, so code that reads content[0].text breaks. Select blocks by their type field.
• Output limits. max_tokens covers thinking plus response text, so limits tuned for no-thinking workloads may cut answers short.
• Beta header. Opus 5.5 fast mode needs no beta header.

Teams on Claude Managed Agents have less to do. The guide says they only need to update the model name.

Why it matters for teams paying for speed

Fast mode is a premium you pay for output speed. BaristaLabs expects it to matter most where someone is waiting on the answer: interactive coding tools, agents a person is watching, or customer-facing chat. Any of that traffic still on Opus 4.8 fast mode at removal starts returning errors.

BaristaLabs' view is that the deadline is the smaller risk, because it's dated and easy to schedule. The larger one is a budget surprise for workloads that ran without thinking on 4.8: a lower rate, but new thinking tokens billed as output. Both risks have the same fix: a short measured test before you switch production traffic.

Recommended first step: rerun one fast workload on Opus 5.5

BaristaLabs recommends picking the single highest-volume request path that sets speed: "fast" on Opus 4.8 and running a sample of it against claude-opus-5-5 with fast mode, at medium effort and then low. Compare three numbers per request: output tokens, total cost at the rates above, and output tokens per second. Check usage.speed to confirm the requests actually ran fast.

If cost per request comes out lower and quality holds, move that path and repeat for the next one. If cost per request comes out higher and a lower effort level hurts quality, consider whether that path needs fast mode at all. Standard-speed Opus 5.5 costs half of its fast rate. For model routing more broadly, the Claude Fable 5 routing policy post covers routing rules for fallback, retention, cost and long-horizon work, and Claude on Google Cloud: Anthropic managed vs Vertex covers the Anthropic managed Google Cloud route, where Opus 5.5 fast mode is also available. For help planning the change, see BaristaLabs' AI consulting services.

Sources and scope

• Claude Platform release notes, Anthropic, entry dated October 30, 2026: deprecation and November 30, 2026 removal of Opus 4.8 fast mode.
• Fast mode, Anthropic documentation, read October 31, 2026: speed, rate limits, caching, Batch and Priority Tier limits, fast mode pricing, model support and beta header.
• Pricing, Anthropic documentation, read October 31, 2026: standard and fast mode rates, platform availability for Opus 5.5 fast mode.
• Claude Opus 5.5 migration guide, Anthropic documentation, read October 31, 2026: adaptive thinking, rejected request settings, response parsing and Managed Agents.

These sources don't publish speed or quality comparisons between Opus 4.8 and Opus 5.5 in fast mode, and the cost arithmetic above is BaristaLabs' calculation from list prices, not a measured result.