Claude API deprecates Opus 4.8 fast mode ahead of November 30 removal, with Opus 5.5 fast mode 20% cheaper

1 hour ago

7 min read

Anthropic has deprecated fast mode for Claude Opus 4.8 on the Claude API, according to an entry dated October 30, 2026 in the Claude Platform release notes. The faster inference configuration will be removed on November 30, 2026, and from that date any request to `claude-opus-4-8` that sets `speed: "fast"` will return an error instead of being served. Anthropic's instruction to developers who want to keep the speed benefit is to migrate to Claude Opus 5.5, the model that now carries fast mode as a generally available feature.

The change is narrow in scope but practical in effect. Claude Opus 4.8 itself is not being retired: the release notes contain no deprecation entry for the model, and its standard-speed pricing remains listed on the official pricing page. What disappears is the premium, faster way of running it. For teams that adopted Opus 4.8 fast mode for interactive coding tools, live agents or latency-sensitive chat, that leaves a thirty-one-day window to move traffic, and the destination is cheaper per token: fast mode on Opus 5.5 is priced at $8 per million input tokens and $40 per million output tokens, against $10 and $50 on Opus 4.8, a reduction of 20% on both sides.

Fast mode, as Anthropic's documentation describes it, is not a separate model. It runs the same weights with a faster inference configuration that delivers up to 2.5x higher output tokens per second than standard speed, with no stated change in intelligence or capabilities. The gain applies to output throughput rather than time to first token, which is why Anthropic says it is most visible with streaming.

··········

WHAT ANTHROPIC ANNOUNCED ON OCTOBER 30.

The release-note entry is short and specific. Anthropic states that it has deprecated fast mode for Claude Opus 4.8, with removal on November 30, 2026; that after removal, requests to `claude-opus-4-8` with `speed: "fast"` will return an error; and that developers who want to continue using fast mode should migrate to Claude Opus 5.5. The fast mode documentation page now marks Opus 4.8 as "Deprecated" in its table of supported models and links directly to the section of the Opus 5.5 migration guide dedicated to moving from Opus 4.8.

The wording on behaviour after removal matters for anyone with fallback logic. When Anthropic removed fast mode for Claude Opus 4.7 on July 24, 2026, the release notes explicitly said that such requests return an error and, unlike Claude Opus 4.6, do not fall back to standard speed. The October 30 entry uses the same formulation for Opus 4.8, so production code that keeps sending `speed: "fast"` to Opus 4.8 after November 30 should expect failed requests, not silently slower ones. Opus 4.8 at standard speed remains available.

This is the third time in 2026 that Anthropic has ended fast mode on an older Opus model. Anthropic deprecated fast mode for Opus 4.6 when Opus 4.8 launched on May 28, removed it for Opus 4.7 in July, and now applies the same pattern to Opus 4.8 roughly five weeks after Claude Opus 5.5 arrived on September 22. The cycle suggests that fast mode is being treated as a capacity-intensive feature that Anthropic concentrates on its newest Opus generation rather than maintaining across the whole catalogue.

··········

FAST MODE PRICES BEFORE AND AFTER THE MIGRATION.

Prices below are as listed on Anthropic's official pricing page and fast mode documentation on October 30, 2026, in USD per million tokens. The official pages publish fast mode input and output prices; they state that prompt caching multipliers apply on top of fast mode pricing but do not print fast mode cache prices, so the cache figures in the table are derived from those documented multipliers (1.25x for five-minute cache writes, 0.1x for cache hits on Opus 4.8 and 0.05x for cache hits on Opus 5.5).

........

Model and speed	Input	Output	5-min cache write	Cache hit	Batch	Status
Opus 4.8, fast mode	$10	$50	$12.50 (derived)	$1.00 (derived)	Not available	Removed Nov 30, 2026
Opus 5.5, fast mode	$8	$40	$10.00 (derived)	$0.40 (derived)	Not available	Available since Oct 15, 2026
Opus 4.8, standard	$5	$25	$6.25	$0.50	$2.50 / $12.50	Remains available
Opus 5.5, standard	$4	$20	$5	$0.20	$2 / $10	Available

........

On both models, fast mode costs exactly twice the standard input and output rate, and Anthropic says the fast mode price applies flat across the full context window, including requests above 200,000 input tokens. Fast mode is not available through the Batch API and cannot be combined with a Priority Tier commitment. Data residency multipliers stack on top: requests pinned to US inference with `inference_geo: "us"` carry a 1.1x multiplier on every token category.

The difference in cache-hit pricing is larger than the 20% headline. Because Opus 5.5 bills cache hits at 0.05x of the input rate rather than 0.1x, a cached token in fast mode costs an estimated $0.40 per million on Opus 5.5 against $1.00 on Opus 4.8. Agentic coding sessions, which resend long conversation prefixes on every turn, are exactly the workloads where that line dominates the input bill.

··········

WHAT A FAST MODE CODING AGENT COSTS BEFORE AND AFTER.

To make the change concrete, consider an interactive coding assistant running in fast mode for one developer team. The assumptions are illustrative: 200 requests per working day, 40,000 input tokens per request, 85% of input tokens served from the prompt cache and the remaining 15% written to the five-minute cache, 2,000 output tokens per request, and 22 working days per month. That gives 6.8 million cached input tokens, 1.2 million cache-write tokens and 0.4 million output tokens per day.

........

Daily cost component	Opus 4.8 fast mode	Opus 5.5 fast mode
Cache hits (6.8M tokens)	$6.80	$2.72
Cache writes (1.2M tokens)	$15.00	$12.00
Output (0.4M tokens)	$20.00	$16.00
Total per day	$41.80	$30.72
Total per month (22 days)	$919.60	$675.84

........

Under these assumptions, the same token volume costs about 26.5% less on Opus 5.5 fast mode, or roughly $244 per month for this single workload. The saving exceeds 20% because of the cheaper cache reads. For comparison, the identical workload at standard speed would cost $20.90 per day on Opus 4.8 and $15.36 per day on Opus 5.5, which shows what the speed premium buys: fast mode on Opus 5.5 costs about twice standard Opus 5.5, but less than fast mode cost on the older model.

The arithmetic holds token counts constant, and that is the main caveat. Opus 5.5 always runs adaptive thinking, and its default effort level is `medium` where Opus 4.8 defaults to `high`. Both differences can change how many output tokens a real request produces, in either direction, so teams should measure their own traffic on Opus 5.5 rather than assuming the savings carry over one-for-one.

··········

WHAT DEVELOPERS NEED TO CHANGE BEFORE NOVEMBER 30.

The minimum change is the model ID: replace `claude-opus-4-8` with `claude-opus-5-5`, which takes no date suffix. On the Messages API, fast mode is still enabled with the same request parameter, as shown in Anthropic's documentation example:

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    speed="fast",
    messages=[
        {"role": "user", "content": "Refactor this module to use dependency injection"}
    ],
)

for block in response.content:
    if block.type == "text":
        print(block.text)

The access model also gets simpler. Fast mode on Opus 4.8 is still a research preview that requires the `fast-mode-2026-02-01` beta header and provisioned access. Since October 15, fast mode on Opus 5.5 is out of research preview: requests no longer need the beta header, organizations no longer need to request access, and Opus 5.5 has a default fast mode rate limit in the Start, Build and Scale usage tiers. Fast mode limits are separate from standard Opus limits and are reported through dedicated `anthropic-fast-input-tokens-*` and `anthropic-fast-output-tokens-*` response headers.

The migration guide lists behavioural differences that deserve more attention than the model string. On Opus 5.5, thinking cannot be disabled: sending `thinking: {"type": "disabled"}` or a fixed `budget_tokens` value returns a 400 error, and a request that omits the field runs with adaptive thinking. Sampling parameters such as `temperature`, `top_p` and `top_k` must be omitted, assistant-turn prefill is rejected, and forced tool choice (`any` or a named `tool`) is no longer accepted. Because thinking blocks can come first in a response, Anthropic recommends selecting content blocks by type rather than by position, and `max_tokens` now has to cover thinking plus the visible answer. Opus 5.5 also lowers the minimum cacheable prompt from 1,024 to 512 tokens, while Priority Tier, which Opus 4.8 supports, is not available on Opus 5.5.

Fallback code should be reviewed as well. Anthropic's documentation recommends catching fast mode rate-limit errors and retrying without `speed: "fast"`, but warns that switching speeds causes a prompt cache miss, because requests at different speeds do not share cached prefixes. A fallback that fires often can therefore erase part of the cache savings described above.

··········

WHERE FAST MODE IS AVAILABLE AND WHAT IT MEANS FOR CLAUDE CODE.

On the Messages API and in Managed Agents, fast mode for Opus 5.5 is available on the Claude API, on Claude Models on Google Cloud (Anthropic managed) since October 22, and on Claude in Microsoft Foundry for Azure-hosted deployments since October 15. The fast mode documentation does not list Amazon Bedrock. Fast mode is eligible for zero data retention, excluding Covered Models, and for HIPAA-eligible use.

In Claude Code, the Opus 4.8 deprecation mostly confirms a shift that already happened in the client. Claude Code's documentation states that Opus 5.5 is the fast mode default in version 2.1.280 and later, after earlier versions defaulted to Opus 5 and, from version 2.1.154 through 2.1.218, to Opus 4.8. Users on Pro, Max, Team and Enterprise plans pay for fast mode from usage credits rather than from their plan's included usage, Team and Enterprise organizations need an Owner to enable it, and the `/fast` command toggles it on and off. Users who pinned Opus 4.8 manually are the ones who should check their model setting before the removal date. For subscription plan changes on the consumer side, Data Studios covered the renaming of Max 20x to Ultra earlier today, and API teams tracking spend can use the Usage and Cost Analytics API, which Anthropic launched in beta on October 29.

··········

WHAT IS ESTABLISHED AND WHAT REMAINS OPEN.

The established facts come from Anthropic's own documentation. Fast mode on Claude Opus 4.8 is deprecated as of October 30, 2026 and will be removed on November 30, 2026; requests to `claude-opus-4-8` with `speed: "fast"` will then fail; the model itself remains available at standard speed and standard prices; and the migration target, Opus 5.5 fast mode, is generally available at $8 and $40 per million input and output tokens, 20% below the $10 and $50 charged for Opus 4.8 fast mode.

Several points are not documented. Anthropic has not said whether fast mode for Claude Opus 5, which remains in research preview at $10 and $50, will follow the same path, although the pattern of the past six months points in that direction. The pricing pages do not print fast mode cache prices explicitly, so the cache figures above rely on the documented multipliers. And the real cost of migrating depends on how Opus 5.5's always-on adaptive thinking and lower default effort change token volumes for each workload. Teams that run Opus 4.8 in fast mode have a clear deadline, a cheaper destination and a short list of request-format changes to test before moving production traffic.

··········

·····

DATA STUDIOS

·····

datastudios.org