Claude Opus 4.8 Fast Mode Ends Nov 30: Migrate to Opus 5.5 or Stay Standard?

Fast mode is a faster inference configuration of the same model. Anthropic’s fast mode page says there is no change to intelligence or capabilities. The speed claim on that page is up to 2.5 times higher output tokens per second than standard speed. The gain is described as output tokens per second, not time to first token. Streaming is compatible, and that gain is most visible there.

You opt in. On the Messages API, set speed to fast. On Managed Agents, set speed on the model object, for example {"id": "claude-opus-5-5", "speed": "fast"}. A successful fast request reports usage.speed as fast. Standard reports standard.

Fast mode has its own rate limit, separate from standard Opus limits. When that limit is exceeded, the API returns HTTP 429 and a retry-after header. The fast mode page also says that exceeding fast mode rate limits or capacity can return 429 or 529. Response headers for the fast pool are anthropic-fast-input-tokens-limit, anthropic-fast-input-tokens-remaining, anthropic-fast-input-tokens-reset, and the matching output headers. The Claude Console rate limits page is where Anthropic tells you to read your own fast limits.

Anthropic defines MTok as million tokens. $5 per MTok means $5 for every million tokens. Fast mode is priced across the full context window, including requests over 200,000 input tokens. There is not a second fast rate for long prompts on this card. Claude 4.6 and later models, except Claude Haiku 5.5, include the 1M token context window at standard per-token rates. Opus 4.8 and Opus 5.5 sit in that group. A 900,000-token request is billed at the same per-token rate as a 9,000-token request, on the standard card. Fast mode keeps one premium rate across that window.

The fast rows match on the fast mode page and the pricing page. Standard and batch rows are the pricing page. Batch is a 50 percent discount on input and output. Fast mode is not available on the Batch API, so one request cannot take both the fast premium and the batch discount.

Cache prices on the standard card matter because agent traffic often rereads a prefix.

Units are USD per million tokens. A 5-minute cache write is 1.25 times the base input price. A 1-hour cache write is 2 times base input. A cache hit on Opus 4.8 is 0.1 times base input, which is the $0.50 row. A cache hit on Opus 5.5 is 0.05 times base input, which is the $0.20 row. The pricing page states that 0.05 times rule for Claude Opus 5.5 and Claude Sonnet 5.5. The general 0.1 times rule is what the Opus 4.8 row follows.

Prompt caching multipliers apply on top of fast mode pricing. The pricing page does not print a separate fast-cache dollar table. Use the multiplier times the fast input price. Data residency also stacks: for Claude 4.6 and later, inference_geo US-only inference is 1.1 times input, output, cache writes, and cache reads on the Claude API, Claude Platform on AWS, and Anthropic-managed Google Cloud. Foundry applies that 1.1 times rule to US Data Zone Standard deployments. Google-managed Google Cloud and Amazon Bedrock publish their own regional prices. This article does not copy those cards.

On the Messages API, Claude Opus 5.5 fast mode is Available on the Claude API, on Claude Models on Google Cloud when Anthropic manages the deployment, and on Claude in Microsoft Foundry when the deployment is hosted on Azure. It left research preview on October 15, 2026. Requests no longer require the fast-mode-2026-02-01 beta header, and requests that still send the header keep working. October 22, 2026 is the release-note date for that Google Cloud path. The October 15 note is the Foundry date.

Managed Agents list Opus 5.5 fast mode as Available on the Claude API. Opus 5.5 has a default fast rate limit on Start, Build, and Scale, with no access request. Evaluation-tier organizations get that limit after they move to one of those tiers.

Claude Opus 4.8 fast mode and Claude Opus 5 fast mode remain in research preview. Both need the fast-mode-2026-02-01 beta header. The fast mode page lists that preview on the Claude API. To use fast mode on Claude Opus 5, Anthropic says to contact your account manager or join the waitlist. The removal note tells people who want to keep fast mode to migrate to Claude Opus 5.5, not to Claude Opus 5.

Fast mode is excluded from the Batch API and from a Priority Tier commitment. The Messages endpoint with speed: "fast" is listed as eligible for zero data retention and HIPAA, with zero data retention applying regardless of speed, excluding covered models. Read that block before you treat fast mode as a new retention regime.

Claude Opus 4.6 is the trap in the error handler. With the beta header, speed: "fast" on Opus 4.6 succeeds, runs at standard speed, reports usage.speed as standard, and is billed at standard rates. On Anthropic-managed Google Cloud, the same Opus 4.6 request returns HTTP 400 even with the header. Opus 4.7 fast mode already returns an error and does not fall back. After November 30, 2026, Opus 4.8 follows that error path. Do not budget an Opus 4.8 fast call as a quiet downgrade to standard.

Keep speed: "fast" when output tokens per second are the constraint you are paying to shrink. The published claim is that speed, up to 2.5 times standard. It is not a time-to-first-token claim, and this article adds no benchmark.

Move the flag to claude-opus-5-5 if you still want it after November 30, 2026. Opus 5.5 fast is $8 and $40 against Opus 4.8 fast at $10 and $50. Drop the preview-header requirement, confirm Start, Build, or Scale, and read usage.speed. standard means you did not buy fast mode.

Drop the flag if the job can finish at standard speed. Opus 4.8 standard remains $5 and $25. Opus 5.5 standard is $4 and $20. Opus 5 standard matches Opus 4.8 at $5 and $25. Staying on Opus 4.8 standard is a real option: the removal hits the fast flag, not the model id. Moving that traffic to Opus 5.5 standard cuts the list rate, and it changes the API. Opus 5.5 launched on September 22, 2026. Thinking cannot be turned off. thinking set to disabled or enabled returns HTTP 400. Omit thinking and set depth with the effort parameter. tool_choice any and tool also return HTTP 400. Use auto with strict tool use. On the Claude API and Google Cloud, computer use requires computer_toolset_20260801, and computer_20251124 returns HTTP 400. On Amazon Bedrock, computer_20251124 keeps working. The pricing table leaves the “any, tool” system-prompt cell blank for Opus 5.5 and lists 286 tokens for auto or none. Opus 4.8 lists 290 and 410. Those counts are extra input tokens when tools are present. They are not a quality score.

Use Batch when the work can wait. Opus 5.5 batch is $2 and $10. Opus 4.8 batch is $2.50 and $12.50. That is a queue, not a live stream. Leave Priority Tier traffic off the fast flag. The November 30 removal does not add a fast row inside that commitment.

A cache hit is a fraction of the input price for the speed you ran. Standard Opus 5.5 hits are $0.20 per million tokens. Standard Opus 4.8 hits are $0.50. The pricing page’s break-even line is about the write, not about your traffic: a 5-minute write at 1.25 times pays off after one hit, and a 1-hour write at 2 times pays off after two hits, if those hits happen.

Fast mode changes the base those multipliers sit on. For a fast Opus 5.5 request, the stacking rule means the fast input price of $8 is the base. A 0.05 times hit is $0.40 per million tokens. A 1.25 times five-minute write is $10. A 2 times one-hour write is $16. For fast Opus 4.8, until removal, the base is $10. A 0.1 times hit is $1. A 1.25 times write is $12.50. A 2 times write is $20. Those six numbers are this article applying the published multipliers to the published fast input prices. They are not a second table on Anthropic’s pricing page. If a bill prints different cache dollars, the invoice wins, and the page to re-check is the pricing page.

Requests at different speeds do not share cached prefixes. Falling back from fast to standard is a prompt cache miss. An agent that retries a 429 by deleting speed pays the standard rate and pays the uncached input on that retry. The prefix you just paid to write at the fast multiplier is not reused at the other speed.

The SDK retries a fast-mode 429 up to 2 times by default. max_retries changes that count. Each retry waits for retry-after. Anthropic says the delay is often short because the fast pool replenishes continuously. Those retries stay on fast mode. They do not create the cross-speed cache miss. They can still hold a user-facing loop.

If you would rather miss the cache than wait, set max_retries to 0 on the fast attempt, catch the rate-limit error, and send again without speed: "fast". That setting also skips retries for other transient errors. The documented example reissues overloaded and connection failures with a separate attempt count. A 500-class error is a reason to retry the same speed. A 429 on the fast pool is the case for a fallback.

Log usage.speed next to the model id. A sum that mixes standard Opus 5.5 at $4/$20 with fast Opus 5.5 at $8/$40 is a blend of two rates.

Opus 4.8’s release note, May 28, 2026, says the minimum cacheable prompt length on that model is 1,024 tokens. A shorter prefix is not a cache write. The same note says Claude Code Max plan users then defaulted to fast mode on Opus 4.8. This November 9 reading of the overview does not show a later line that cancels that default. If Claude Code is still sending speed: "fast" on claude-opus-4-8, that traffic is inside the November 30 removal. Check the request, not the plan name.

With at least one tool, Opus 5.5 adds 286 system-prompt tokens for tool_choice auto or none. Opus 4.8 adds 290, or 410 for any or tool. Client-side tools are priced like any other request. Server-side tools can add their own charges. Web search on the Claude API stays $10 per 1,000 searches. The tokens around the search follow the speed you ran.

Every figure below is illustrative. It is not a customer bill. It uses public list rates checked on November 9, 2026, and a usage assumption you can see. Taxes, discounts, data-residency multipliers, and failed calls are outside the formula unless the line says they are in it.

Assumption A: 2 million uncached input tokens and 0.5 million output tokens, one request shape, no cache, no tools, no inference_geo. Formula: millions of input tokens times the input rate, plus millions of output tokens times the output rate.

Opus 4.8 fast, while it still exists: 2 × $10 + 0.5 × $50 = $20 + $25 = $45.

Opus 5.5 fast: 2 × $8 + 0.5 × $40 = $16 + $20 = $36.

Opus 4.8 standard: 2 × $5 + 0.5 × $25 = $10 + $12.50 = $22.50.

Opus 5.5 standard: 2 × $4 + 0.5 × $20 = $8 + $10 = $18.

Opus 5.5 batch, if the job can wait: 2 × $2 + 0.5 × $10 = $4 + $5 = $9.

On this mix, Opus 4.8 fast at $45 becomes Opus 5.5 fast at $36, Opus 4.8 standard at $22.50, Opus 5.5 standard at $18, or Opus 5.5 batch at $9. The order is a property of these token counts. A job that is almost all cached input will not sort the same way. Recompute before you paste $45 into a forecast.

Assumption B: the same 0.5 million output tokens, but 1.5 million of the input tokens are cache hits and 0.5 million are uncached. No write on this call. Standard Opus 5.5 uses the published $0.20 hit price. Fast Opus 5.5 uses the derived $0.40 hit price from 0.05 × $8.

Standard: 0.5 × $4 + 1.5 × $0.20 + 0.5 × $20 = $2 + $0.30 + $10 = $12.30.

Fast, derived hit price: 0.5 × $8 + 1.5 × $0.40 + 0.5 × $40 = $4 + $0.60 + $20 = $24.60.

The pricing page prices output as output. It does not publish a separate Opus thinking-token rate. A cross-speed fallback is a cache miss plus a second request at the standard rate. The page does not say a 429 is free of input charges, so read the invoice before you treat the failed attempt as zero. If inference_geo selects US-only inference, multiply the token result by 1.1. Assumption A on Opus 5.5 fast becomes $36 × 1.1 = $39.60, still illustrative.

1. Find every claude-opus-4-8 call with speed: "fast", including Managed Agents and any Claude Code path still defaulting there.
2. Split the list into must stay fast, can run at standard speed, and can wait for Batch.
3. On the fast pile, set claude-opus-5-5 and keep speed: "fast". The fast-mode-2026-02-01 header is no longer required. Requests that still send it keep working.
4. Confirm Start, Build, or Scale. Evaluation tier traffic does not get the default Opus 5.5 fast limit.
5. On the standard pile, delete speed. Keep claude-opus-4-8 at $5/$25 or move to claude-opus-5-5 at $4/$20, and apply the thinking and tool-choice constraints.
6. On the batch pile, remove speed and use the Batch API. Budget $2/$10 or $2.50/$12.50. A fast flag on a batch request is not a discount.
7. Rebuild cache breakpoints per speed. A fast write is not a standard hit.
8. Log usage.speed, model id, cache reads, and cache writes on every attempt.
9. On 429, wait for retry-after or fall back once. A fallback on every agent step pays the cache miss again.
10. On 400, read the body. Waiting will not fix tool_choice any or tool, a disabled thinking block, or computer_20251124 on Opus 5.5 outside Bedrock.
11. Watch the fast rate-limit headers. The standard pool can look idle while the fast pool is already returning 429.
12. After one day, recompute assumption A with your token counts. Keep the list rates. Replace the 2 million and the 0.5 million.

Opus 5 fast mode stays a preview, with a waitlist and the beta header, at $10 and $50. The removal note’s target is Opus 5.5. Stay on Opus 5 fast only if you are already in that preview and you mean to.

When does Claude Opus 4.8 fast mode stop working?

November 30, 2026. Anthropic deprecated it on October 30, 2026. After removal, requests to claude-opus-4-8 with speed: "fast" return an error. Standard-speed requests to that model id are a separate row on the pricing page.

Did Claude Opus 5.5 fast mode leave research preview?

Yes. On October 15, 2026 it left research preview, access requests stopped being required, and the beta header stopped being required. The fast mode page lists the status as Available. Evaluation-tier organizations still move to Start, Build, or Scale before the default fast limit applies.

What are the fast and standard prices?

Per million tokens, in US dollars: Opus 4.8 fast is $10 input and $50 output until removal. Opus 5.5 fast is $8 and $40. Opus 4.8 standard is $5 and $25. Opus 5.5 standard is $4 and $20. Fast prices apply across the full context window, including prompts over 200,000 input tokens. Checked November 9, 2026.

Do fast and standard requests share a prompt cache?

No. The fast mode page says requests at different speeds do not share cached prefixes, and a fallback from fast to standard is a cache miss. Cache multipliers apply on top of whichever speed you actually ran.

What should I log when a fast call fails?

Log the HTTP status, retry-after when it is present, the model id, and usage.speed when a response exists. 429 and 529 are the capacity and rate-limit failures named on the fast mode page. 400 is the failure for an unsupported combination, including Opus 5.5 thinking and tool-choice constraints in the September 22, 2026 release note. Opus 4.6’s old behavior, succeeding at standard speed, is not the Opus 4.8 behavior after November 30.