# Opus 4.8 fast mode ends Nov 30: find every caller and re-price

URL: https://infro.io/blog/opus-4-8-fast-mode-ends-nov-30-find-every-caller-and-re-price
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-11-08
Last modified: 2026-11-08
Last reviewed: 2026-11-08 (next review due 2026-12-08)
Topic: Cost and FinOps
Format: Checklist

> Fast mode on Claude Opus 4.8 is removed on November 30, 2026, and requests that still set speed "fast" will return an error rather than run slower [^anthropic-relnotes]. Find the callers now, drop or move the flag, and re-price: list rates halve, but output speed falls by up to 2.5x [^anthropic-fast-mode].

Key takeaways:
- After November 30, 2026, a request to claude-opus-4-8 with speed "fast" returns an error; it does not quietly run at standard speed.
- Standard Opus 4.8 lists at half the fast-mode rate, but fast mode's up-to-2.5x output speed goes with it, so total latency SLOs can break even as the bill falls.
- Fast mode speeds up output tokens per second, not time to first token, so a TTFT objective should be unaffected and a total-duration objective is the one at risk.
- Moving to Opus 5.5 keeps fast mode but brings always-on adaptive thinking and rejected request settings, so it is a migration, not a model-id swap.
- Switching speed misses the prompt cache, so expect cache writes on the first requests after cutover.

## What changes on November 30

On October 30, 2026 Anthropic deprecated fast mode for Claude Opus 4.8, with removal on November 30. After removal, requests to `claude-opus-4-8` with `speed: "fast"` return an error, and Anthropic points callers who want to keep fast mode to Claude Opus 5.5.[1] That is 22 days from today, November 8.

Two details shape the work. First, it is a hard failure, not a silent downgrade: a caller you miss turns into an error rate. The fast mode page does describe a fallback pattern, but it catches rate-limit errors (429) and retries without `speed`; a removed feature is a different error, and the release note says only "an error".[2] Second, the model is not going anywhere. Anthropic lists `claude-opus-4-8` as active, with retirement not sooner than May 28, 2027.[3] Only the flag is removed, so the cheapest fix is often to stop sending it.

Fast mode on Opus 4.8 was a research-preview feature, and Anthropic's notes say to keep sending the beta header with fast requests to that model.[1] So there are two strings to search for: the `speed` field and the `fast-mode-2026-02-01` header.

## Step 1: find every caller

Do the code search and the traffic search, because each misses things the other catches. Code search misses feature flags, config files and third-party agents; traffic search misses callers that only run monthly.

- [ ] Grep every repository, infrastructure config and notebook for `speed`, `"fast"` and `fast-mode-2026-02-01`.
- [ ] Search for the model id `claude-opus-4-8` in environment variables, prompt-management tools and route definitions, not only in code.
- [ ] For Claude Managed Agents, check agent definitions: speed is set on the agent's model object, for example `{"id": "claude-opus-5-5", "speed": "fast"}`.[2]
- [ ] Log `usage.speed` from every Opus 4.8 response for a week. The response reports `"fast"` or `"standard"`, which tells you which calls really ran fast.[2]
- [ ] Watch the fast-mode rate-limit response headers (`anthropic-fast-output-tokens-limit` and related) in your logs; their presence marks a fast request.[2]
- [ ] Pull a usage export grouped by API key and model from the Claude Console, which Anthropic documents for auditing deprecated usage, and match keys to owners.[3]
- [ ] If you call Claude through a gateway, group Opus 4.8 requests by key, project and feature, so every owner of an Opus 4.8 key is asked.
- [ ] Ask the owners of scheduled jobs and evaluation harnesses, which rarely appear in a one-week window.

A single shim makes the cutover safe while the owners fix their code. It strips the flag for this model after a switch date and logs what it removed:

```python
import logging
from datetime import datetime, timezone

CUTOFF = datetime(2026, 11, 30, tzinfo=timezone.utc)
log = logging.getLogger("fast-shim")

def create(client, **params):
    if params.get("model") == "claude-opus-4-8" and params.get("speed") == "fast":
        if datetime.now(timezone.utc) >= CUTOFF:
            params.pop("speed")
            # drop the research-preview beta header too if you send it
            log.warning("removed speed=fast for opus-4-8 caller=%s", params.get("metadata"))
    return client.messages.create(**params)
```

## Step 2: decide per caller

Not every fast caller needs the same answer. Sort each one by what its latency target is made of.

| Caller | What it needs | First option to test |
| --- | --- | --- |
| Interactive chat, streamed | Fast token stream | Opus 5.5 with `speed: "fast"` |
| Agent loop, many turns | Short total run time | Opus 5.5 standard or fast; fewer turns |
| Background job nobody waits on | Throughput, not speed | Drop the flag, stay on Opus 4.8 |
| Short answers, tight SLO | Low total time | Sonnet 5.5 or Haiku 5.5 |
| Batch | n/a | Fast mode is not available in the Batch API[2] |

The background-job row is the most common saving: a caller that set fast mode by habit and has no user waiting pays double for nothing.

## Step 3: re-price

Fast mode is a multiplier on standard rates across the full context window.[2] On Opus 4.8 and Opus 5 it lists at $10 input and $50 output per million tokens; on Opus 5.5 at $8 and $40.[2] Standard Opus 4.8 lists at $5 and $25, and standard Opus 5.5 at $4 and $20, with cache reads at $0.50 and $0.20.[4] Anthropic's pricing page states the Opus 5.5 fast mode as 2x standard pricing.[5] Sonnet 5.5 lists at $2 and $10, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens.[4] These are list rates; if your contract carries different rates, apply them to the same arithmetic.

> **Example: Illustrative monthly re-price**
> Illustrative figures, not measurements: a caller sends 200M input and 40M output tokens a month with fast mode on Opus 4.8.
> 
> | Option | Input | Output | Monthly total |
> | --- | --- | --- | --- |
> | Opus 4.8 fast (today) | $2,000 | $2,000 | $4,000 |
> | Opus 4.8 standard | $1,000 | $1,000 | $2,000 |
> | Opus 5.5 standard | $800 | $800 | $1,600 |
> | Opus 5.5 fast | $1,600 | $1,600 | $3,200 |
> | Sonnet 5.5 | $400 | $400 | $800 |
> 
> Now assume Opus 5.5's always-on thinking lifts output by 30% (an assumption for the example): output becomes 52M tokens, so Opus 5.5 standard is $800 + $1,040 = $1,840 and Opus 5.5 fast is $3,680.

The last line matters. On Opus 5.5 every request runs with adaptive thinking, thinking tokens bill as output tokens, and workloads that ran without a thinking field on Opus 4.8 can produce more output tokens per request.[6] Re-baseline cost at your chosen effort level, not just from the rate card.

Also plan for cache: requests at different speeds do not share cached prefixes, so moving off fast mode causes a cache miss.[2] Expect the first requests after cutover to pay cache writes, which are 1.25x the base input price for the five-minute duration.[4]

## Step 4: what happens to latency objectives

Fast mode delivers up to 2.5x higher output tokens per second, and Anthropic says the benefit is in output speed, not time to first token.[2] That splits your objectives in two:

- **Time to first token:** should not move when you drop the flag, since fast mode was not what made it quick.
- **Total response time and tokens per second:** can lengthen by up to 2.5x for the generation phase, the biggest effect on long, streamed answers.

> **Example: Illustrative latency arithmetic**
> Illustrative only: if standard speed streams at 50 tokens per second and fast mode at 125, a 1,000-token answer takes about 20 seconds standard versus 8 seconds fast. A p95 objective of 10 seconds for completion would fail on that caller at standard speed, even though the bill halved.

Rate limits change too. Fast mode has a dedicated rate limit separate from standard Opus limits, and exceeding it returns a 429.[2] Callers moved to standard speed fall under the standard Opus limits instead, so check headroom for the combined volume. If any caller is on a Priority Tier commitment, note that fast mode is unavailable with one, and Priority Tier is not supported on Opus 5.5.[2][6]

## Step 5: options to test

Run each option on a replay of your own traffic and score cost per successful task, not rate-card price; the benchmark claims below are Anthropic's, not yours.

- [ ] **Opus 4.8 at standard speed.** Same model, half the list rate, slower output. Test the total-time objective only.
- [ ] **Opus 5.5, standard and fast.** Keeps fast mode and lists lower, but the migration guide requires changes: no `thinking: disabled`, effort set explicitly (default medium), no forced `tool_choice` of `any` or a named tool, no sampling parameters, and no prefilled assistant turn.[6] Read content blocks by type, because responses can start with thinking blocks.
- [ ] **Sonnet 5.5.** Anthropic says it runs 30%+ faster than Sonnet 5 and costs up to 30% less for most work; that comparison is against Sonnet 5, not Opus.[7] Test it on tasks where Opus-level depth was never the constraint.
- [ ] **Haiku 5.5.** Described by Anthropic as its fastest small model, built for high-volume work such as summarization and subagents.[7] The right test for short, latency-bound calls.
- [ ] **Another vendor's fast tier.** Compare against [service tiers across vendors](https://infro.io/blog/llm-service-tiers-compared-2026) and the [ultrafast agent-loop analysis](https://infro.io/blog/ultrafast-mode-in-an-agent-loop-which-steps-earn-the-6x-price), and check equivalence with the method in your fallback tests.

## A dated plan to November 30

1. **Nov 8–14:** run the code search, turn on `usage.speed` logging, and list every Opus 4.8 key and its owner.
2. **Nov 15–19:** classify each caller with the table above and set a latency and cost target for each.
3. **Nov 20–25:** replay traffic on the chosen options in staging; ship the shim as a safety net.
4. **Nov 26–29:** cut over callers by owner, starting with background jobs, and watch error rate and p95 per caller.
5. **Dec 1:** confirm no request carries `speed: "fast"` for `claude-opus-4-8`, and re-forecast December from the new mix.

## Where a gateway helps

If your Claude traffic goes through a gateway, usage per request already records model, latency and estimated provider cost, and can be grouped by key, project, feature or tag, which narrows the owner list in Step 1 and gives you p50/p95 per caller to compare before and after.[8] Routing rules in INFRO can send a matching feature or key to a different target, which lets you move one caller to Opus 5.5 or Sonnet 5.5 without a deploy. Enter your contract rates under negotiated rates so the re-price uses what your provider bills. Confirm in staging whether your path forwards the `speed` field before relying on it.

## Questions

### Does fast mode change the answers I get from Opus 4.8?

Anthropic describes fast mode as the same model on a faster inference configuration, with no change to intelligence or capabilities [2]. Dropping the flag should therefore change speed and price, not answer quality, though thinking and token counts are what to compare if you also change models.

### Can I use fast mode in the Batch API instead?

No. Anthropic's fast mode page lists the Batch API as unsupported, and fast mode is also unavailable with a Priority Tier commitment [2].

## Sources

1. Claude Platform release notes — Anthropic. https://platform.claude.com/docs/en/release-notes/overview (accessed 2026-11-08)
2. Fast mode — Anthropic. https://platform.claude.com/docs/en/build-with-claude/fast-mode (accessed 2026-11-08)
3. Model deprecations — Anthropic. https://platform.claude.com/docs/en/about-claude/model-deprecations (accessed 2026-11-08)
4. Pricing — Anthropic. https://platform.claude.com/docs/en/about-claude/pricing (accessed 2026-11-08)
5. Plans & pricing — Anthropic. https://claude.com/pricing (accessed 2026-11-08)
6. Claude Opus 5.5 migration guide — Anthropic. https://platform.claude.com/docs/en/models/opus-5-5/migration-guide (accessed 2026-11-08)
7. Newsroom — Anthropic. https://www.anthropic.com/news (accessed 2026-11-08)
8. Usage analytics — INFRO. https://infro.io/docs/observability/analytics (accessed 2026-11-08)
