Claude Opus 4.8 Fast Mode Ends November 30: How to Migrate to Opus 5.5 (2026)

On October 30, 2026, Anthropic added a line to the Claude Platform release notes that's easy to scroll past: fast mode for Claude Opus 4.8 is deprecated, with removal on November 30, 2026. After that date, any request to claude-opus-4-8 that sets speed: "fast" returns an error. Not a slowdown. An error.

If you've got a streaming UI, a voice agent, or a coding assistant tuned around that faster output, you have 30 days. The official migration target is Claude Opus 5.5, which is cheaper per token and has fast mode out of preview. But it isn't a drop-in swap, and a few of the differences can quietly raise your bill.

What Actually Changed (and What Didn't)

Only fast mode for Opus 4.8 is going away. Anthropic still lists Opus 4.8 as a legacy model that remains available, so standard-speed calls to claude-opus-4-8 aren't covered by this notice. If you never set speed, nothing breaks on November 30.

Fast mode on Claude Opus 5 stays gated behind a waitlist or an account manager, so Opus 5.5 is the only Opus model where you can simply turn it on. And the speedup is on output tokens per second, not time to first token.

Pricing: What Opus 5.5 Fast Mode Costs

Fast mode is a flat 2x multiplier on standard rates, applied across the full 1M-token context window. Prompt caching multipliers stack on top. All figures below are USD per 1 million tokens, from Anthropic's pricing and models pages as of October 31, 2026.

Here's the math on a workload of 40 million input tokens and 8 million output tokens per month, all on fast mode. Opus 4.8 fast: 40 × $10 + 8 × $50 = $800. Opus 5.5 fast: 40 × $8 + 8 × $40 = $640. That's 20% less.

Now the catch. Thinking tokens on Opus 5.5 are billed as output. If your old setup ran without thinking and the new one generates 50% more output tokens (8 million becomes 12 million), the Opus 5.5 fast bill climbs to $800 and your savings vanish. Measure output tokens per request first.

What Breaks When You Swap the Model ID

Changing the string to claude-opus-5-5 is the easy part. Anthropic's migration guide lists the things that bite:

• Thinking can't be turned off. thinking: {"type": "disabled"} and "enabled" with a budget both return a 400. Drop the field and set effort instead. Default effort is medium, and the supported levels are low, medium, high, xhigh and max.
• Sampling parameters are gone. Remove temperature, top_p and top_k from payloads.
• max_tokens now covers thinking plus the answer. At xhigh or max effort, start at 64k.
• Forced tool use fails. tool_choice types any and tool return a 400.
• Prefill is rejected.
• No Priority Tier support.

One more thing worth a unit test: the response now always contains thinking blocks before the text block. Code that reads content[0].text will break. Find the text block by its type.

Working Code: Fast Mode on Opus 5.5

curl, with no beta header needed on Opus 5.5:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 4096,
    "speed": "fast",
    "output_config": {"effort": "medium"},
    "messages": [{"role": "user", "content": "Refactor this module to use dependency injection"}]
  }'

Fast mode has its own rate limit. Exceed it and you get a 429 with a retry-after header (capacity problems can surface as a 529), and the SDK retries twice by default. To degrade to standard speed instead of waiting, use this in Python:

import anthropic

client = anthropic.Anthropic()

def ask(prompt: str, fast: bool = True):
    params = dict(
        model="claude-opus-5-5",
        max_tokens=4096,
        output_config={"effort": "medium"},
        messages=[{"role": "user", "content": prompt}],
    )
    if fast:
        params["speed"] = "fast"
    try:
        # max_retries=0 so a fast-mode 429 surfaces immediately
        return client.with_options(max_retries=0).messages.create(**params)
    except anthropic.RateLimitError:
        if not fast:
            raise
        params.pop("speed")  # standard speed; cached prefixes won't be shared
        return client.messages.create(**params)

resp = ask("Refactor this module to use dependency injection")
print(resp.usage.speed)  # "fast" or "standard"

Setting max_retries=0 also disables 5xx retries on the fast attempt, so production code needs a longer helper, like the one in Anthropic's fast mode docs. Node.js, with the text-block lookup the new response shape needs:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const msg = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 4096,
  speed: "fast",
  output_config: { effort: "medium" },
  messages: [{ role: "user", content: "Summarize this stack trace" }],
});

const text = msg.content.find((b) => b.type === "text");
console.log(msg.usage.speed, text?.text); // "fast" or "standard"

Log usage.speed on every call so you know which speed you actually paid for.

Is Fast Mode Worth Keeping?

My take: keep it only where a human watches tokens stream, like chat panels, inline code edits and voice. Background agents and nightly jobs pay double for speed nobody sees, and can't use fast mode in the Batch API anyway.

Before paying $8/$40, run Claude Sonnet 5.5 at standard speed on the same prompts. At $2/$10 it costs a quarter of Opus 5.5 fast mode. If your evals pass, the "fast Opus" requirement may disappear. A/B tests like this are easier when several models sit behind one OpenAI-compatible endpoint, which is how a lot of teams use KissAPI: change the model string, rerun the suite, compare. Just confirm that any speed-specific parameter you depend on passes through whatever gateway you use.

Opus 5.5 Fast vs the Alternatives

A 30-Day Rollout Plan

1. By November 7: inventory. Grep your codebase and gateway configs for speed, fast-mode-2026-02-01, and claude-opus-4-8.
2. By November 14: shadow test. Replay real prompts against Opus 5.5 at medium effort. Compare output tokens, latency, and eval scores.
3. By November 21: ramp. Move 10% of fast-mode traffic, then 50%, then everything. Watch your 429 rate and cache hit ratio.
4. November 22 to 29: buffer. Keep a feature flag that drops speed entirely if something misbehaves.

Anthropic's Usage and Cost Analytics API (beta since October 29, 2026) filters by model, workspace and API key, which makes before-and-after spend easy to compare. To size the bill ahead of time, plug your token volumes into the API cost calculator and measure real prompts with the token counter.

Frequently Asked Questions

When does fast mode stop working for Claude Opus 4.8?

Anthropic deprecated fast mode for Claude Opus 4.8 on October 30, 2026 and will remove it on November 30, 2026. After removal, requests to claude-opus-4-8 with speed set to fast return an error. The deprecation notice covers fast mode only, not standard-speed requests to claude-opus-4-8.

How much does fast mode cost on Claude Opus 5.5?

Fast mode on Claude Opus 5.5 costs $8 per million input tokens and $40 per million output tokens, which is 2x the standard $4 and $20 rates. Prompt caching and data residency multipliers apply on top of those prices, and fast mode is not available through the Batch API.

Do I need a beta header or special access to use fast mode on Claude Opus 5.5?

No. Fast mode is out of research preview for Claude Opus 5.5, requests no longer need the fast-mode-2026-02-01 beta header, and the model has a default fast mode rate limit in the Start, Build, and Scale usage tiers. Requests that still send the header keep working.