Industry News
Fast mode on Claude Opus 4.8 stops working on November 30, 2026. Opus 5.5 fast mode is 20% cheaper at $8/$40 per 1M tokens, but always-on thinking and API breaking changes decide the real bill.
Some links are partner links: if you subscribe through them, we may earn a commission, at no extra cost to you. The crowd verdicts stay independent.
Opus 4.8 fast mode is deprecated: Anthropic announced it on October 30, 2026, and on November 30, 2026 every request to claude-opus-4-8 that still sends speed: "fast" will return an error. The only supported way to keep fast output is Claude Opus 5.5, which is cheaper per token in fast mode but changes enough of the API to break code written for 4.8.
Move to Opus 5.5 before November 30
Fast mode on Claude Opus 4.8 stops working on November 30, 2026. Claude Opus 5.5 is the recommended replacement: its fast mode costs $8 / $40 per million input / output tokens versus $10 / $50 on Opus 4.8, a 20% cut, and it is generally available without a beta header since October 15.
What the Opus 4.8 fast mode deprecation changes, date by date #
The Claude API release notes set out a short timeline. Teams still running Opus 4.8 at fast speed have 26 days from the date of this article.
| Date (2026) | Event |
|---|---|
| September 22 | Claude Opus 5.5 launches at $4 / $20 per 1M tokens, fast mode in research preview |
| October 15 | Fast mode on Opus 5.5 leaves research preview: no beta header, default rate limit in the Start, Build and Scale tiers |
| October 30 | Fast mode on Opus 4.8 is deprecated |
| November 30 | Removal: claude-opus-4-8 withspeed: "fast" returns an error |
Two details matter. First, the Opus 4.8 model itself is not retired: standard-speed requests keep working. Only the fast tier goes away. Second, Anthropic does not silently downgrade the request to standard speed. The call fails, so any production path that hardcodes speed: "fast" on 4.8 breaks on December 1 if nobody touches it.
Claude Opus 5.5 pricing vs Opus 4.8 and Opus 5 #
Three Opus models currently accept fast mode. Only one of them is a realistic landing spot. The prices below come from the arena catalog (verified against Anthropic's documentation), and the crowd score is the share of GLAD-AI-TOR voters who recommend the model.
| Model | Standard (in / out per 1M) | Fast mode (in / out per 1M) | Fast mode status | Crowd score |
|---|---|---|---|---|
| Claude Opus 4.8 | $5 / $25 | $10 / $50 | Deprecated, removed Nov 30 | 57% (3 votes) |
| Claude Opus 5 | $5 / $25 | $10 / $50 | Research preview, waitlist | 63% (4 votes) |
| Claude Opus 5.5 | $4 / $20 | $8 / $40 | Generally available | No votes yet |
Opus 5 looks like a softer jump, but its fast mode still requires the fast-mode-2026-02-01 beta header and access through an account manager or a waitlist, according to the fast mode documentation. It costs the same as 4.8, and the catalog already lists it as superseded by Opus 5.5. Moving there buys time, not savings.
Opus 5.5 is the opposite case: 20% cheaper at both speeds, cache reads at $0.20 per million tokens versus $0.50 on Opus 4.8, and an arena verdict that calls it "the new value king for frontier agentic work". Its catalog pros include 66.4% on Terminal-Bench 4.0 and more than 30% faster output than Opus 5 at standard speed. Fast mode also travels further: Opus 4.8 fast was Claude API only, while Opus 5.5 fast runs on the Claude API, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry.
Claude Opus 5.5
Anthropic's September 2026 flagship: Fable-level scores at 40% lower cost, $4/$20 pricing, #1 on Artificial Analysis
Partner link. The crowd verdicts stay independent.
What moving to Opus 5.5 really costs: the thinking tax #
The sticker price drops 20%. The bill does not necessarily follow. On Opus 4.8, a request without a thinking field runs without thinking. On Opus 5.5, adaptive thinking is always on and cannot be disabled, and thinking tokens are billed as output tokens even when their text is not returned.
A worked example for a fast-mode workload of 10M input and 2M output tokens per month:
| Scenario | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Opus 4.8 fast, no thinking | 10M x $10 = $100 | 2M x $50 = $100 | $200 |
| Opus 5.5 fast, same output volume | 10M x $8 = $80 | 2M x $40 = $80 | $160 |
| Opus 5.5 fast, output +50% from thinking | 10M x $8 = $80 | 3M x $40 = $120 | $200 |
On this mix, the migration saves $40 a month as long as thinking adds less than 50% to output volume. On output alone, the cheaper rate absorbs up to 25% more output tokens before the bill matches 4.8. The lever is the effort parameter: Opus 5.5 defaults to medium, where Opus 4.8 defaulted to high, and dropping to low on simple calls is the documented way to spend fewer thinking tokens. The catalog's own cons flag the risk: Opus 5.5 is "still verbose", and its high output counts "can exceed headline savings".
Two cost traps sit outside the token price. Fast mode is not available with the Batch API (Opus 4.8 batch pricing is $2.50 / $12.50, so batch jobs never belonged in fast mode anyway). And Priority Tier is not supported on Opus 5.5, so organizations with a Priority Tier commitment on 4.8 need to plan capacity separately.
Opus 5.5 migration checklist for fast mode users #
The Opus 5.5 migration guide lists every change. For a team coming from Opus 4.8 fast, these are the ones that return a 400 error or change behavior:
- Model ID. Replace
claude-opus-4-8withclaude-opus-5-5. Thefast-mode-2026-02-01beta header is no longer needed; requests that still send it keep working. - Thinking. Remove
thinking: {"type": "disabled"}andthinking: {"type": "enabled", "budget_tokens": N}. Both now return a 400. Omit the field and seteffortinstead. - Effort. Set it explicitly. A request that omits it now runs at
medium, nothigh. - Forced tool use.
tool_choicetypesanyandtoolreturn a 400, including on token counting. Useautowith strict tool use or structured outputs. - Sampling and prefill. Omit
temperature,top_pandtop_k, and do not endmessageswith a prefilled assistant turn. - Computer use. Declare the
computer_toolset_20260801toolset;computer_20251124returns a 400 on the Claude API. - Response parsing. Responses can start with
thinkingblocks, so code that readscontent[0].textbreaks. Select blocks bytype, and pass thinking blocks back unmodified in tool-use loops. - max_tokens. It covers thinking plus text. At
xhighormaxeffort, Anthropic suggests starting at 64k. - Fallback. Keep a retry without
speed: "fast"for 429 errors, knowing that switching speed causes a prompt cache miss.
The minimal request change looks like this:
{
"model": "claude-opus-5-5",
"max_tokens": 4096,
"speed": "fast",
"output_config": { "effort": "low" },
"messages": [{ "role": "user", "content": "..." }]
}
One side benefit: the minimum cacheable prompt drops from 1,024 tokens on Opus 4.8 to 512 on Opus 5.5, so short system prompts that never cached before now can.
Claude Opus 5.5
Anthropic's September 2026 flagship: Fable-level scores at 40% lower cost, $4/$20 pricing, #1 on Artificial Analysis
Partner link. The crowd verdicts stay independent.
Which teams can skip fast mode instead #
Not every workload needs to move. Fast mode delivers up to 2.5x output tokens per second, but it does not improve time to first token, and it doubles the per-token price on Opus 4.8. Background agents, nightly jobs and anything that is not streamed to a waiting user can drop speed: "fast" and stay on Claude Opus 4.8 at $5 / $25, with zero code change beyond the flag. That halves the bill immediately and defers the Opus 5.5 migration to a planned sprint.
The catch: Opus 4.8 is already two generations behind in the catalog, and its 57% crowd score sits below Opus 5 (63%). Staying is a holding pattern, not a strategy. Teams weighing a cheaper tier entirely can also read the earlier arena analysis of when Claude Sonnet beats Opus on cost.
The verdict #
- Deadline: fast mode on Opus 4.8 fails from November 30, 2026. Search the codebase for
speedandclaude-opus-4-8now. - Price: Opus 5.5 fast mode costs $8 / $40 per 1M tokens, 20% below the $10 / $50 charged on Opus 4.8 and Opus 5.
- Real cost: savings hold as long as always-on thinking adds less than 25% to output-heavy workloads; tune
effortbefore judging the bill. - Breaking changes: thinking cannot be disabled, forced
tool_choicereturns a 400, sampling parameters and prefill are rejected. - Escape hatch: latency-tolerant jobs can drop fast mode and stay on Opus 4.8 at $5 / $25 while the migration is tested.
For the full head-to-head, see the Claude Opus 4.8 vs Claude Opus 5.5 duel, the Claude Opus 5.5 profile and the arena's ranking of the best LLM models.
Keep exploring
Every claim above is backed by the arena's live data: crowd votes, verified pricing, honest pros & cons.
More from the arena journal #
Industry News
Claude Sonnet 5 vs Opus 4.8: When the $2 Model Beats the $25 One
Sonnet 5 costs 60% less than Opus 4.8 but matches it on knowledge work. Per-task cost analysis reveals when each model wins.
Aug 19, 2026 · 4 min read
Industry News
DeepSeek V4 vs Claude Opus 4.8: Can the $0.87 Model Compete in Agentic Coding?
DeepSeek V4 costs 28.7x less than Claude Opus 4.8 per output token. Both score 57% crowd approval, but the real tradeoff lies in tool-call reliability vs raw price. Here is what the data shows.
Jul 27, 2026 · 4 min read
Industry News
ElevenLabs Signs With Universal Music Group: What the Deal Changes (and Doesn't) for Eleven Music Users
The ElevenLabs Universal Music Group deal funds a separate fan platform. Eleven Music stays at $6/month with the same terms. What buyers actually get today.
Oct 30, 2026 · 6 min read