Claude Opus 5.5 Fast Mode Is Out of Preview: What the 2x Rate Card Buys You

Fast mode for Claude Opus 5.5 left research preview on October 15, 2026. The queue is gone: there is no waitlist to join, no access request to file, and no beta header to attach. Fast mode has become a normal, self-serve part of the Claude API, of Claude Code, and of Claude in Microsoft Foundry for Azure-hosted deployments. The model behind it did not change. What changed is that the gate came down, and the price tag behind the gate is now the only thing standing between you and it — $8 per million input tokens and $40 per million output tokens, against $4 and $20 for standard speed. Exactly double, in both directions.

That is the whole trade. Claude Opus 5.5 shipped on September 22, 2026 at $4 and $20 per million tokens, a 20 percent cut against Claude Opus 5, with cache reads down 60 percent to $0.20 per million. Fast mode was already available then, but as a research preview you had to be admitted to. As of October 15 it is a product you can switch on yourself, with a published rate limit in the ordinary usage tiers — and that shift from admittance to access is what makes the rate card worth reading properly rather than skimming.

What actually changed on October 15

Three things, and only three, and it matters which of them are true of which model:

• Access — fast mode for Claude Opus 5.5 no longer needs a request or an approval. It is on by default in the Start, Build and Scale usage tiers. Organisations still in the Evaluation tier can use it once they move up a tier.

• Protocol — the fast-mode-2026-02-01 beta header is no longer required for requests to Claude Opus 5.5. Requests that still send it keep working, so nothing in an existing integration needs to be removed this week.

• Surfaces — fast mode for Claude Opus 5.5 arrived on Claude in Microsoft Foundry for Azure-hosted deployments on the same day. It is not available on Claude Platform on AWS, on Amazon Bedrock or on Google Cloud, and it is not available on Foundry deployments hosted by Anthropic rather than by Azure.

One thing explicitly did not change. Fast mode for Claude Opus 5 and for Claude Opus 4.8 is still in research preview, and Anthropic's own instruction is that you keep sending the beta header for those two models. If you have code that branches on model ID to decide whether to attach the header, the October 15 change makes Claude Opus 5.5 the one branch that no longer needs it — which is the sort of asymmetry that quietly breaks a shared client if nobody updates the condition.

Same model, and that is the entire appeal

Fast mode is not a second model and it is not a distilled variant. Anthropic is explicit about this: it is the same weights, the same behaviour, the same intelligence, running under a faster inference configuration. Nothing about the answer changes. What changes is how quickly the answer is written out.

The specific number is up to 2.5 times the output tokens per second, and the specific caveat is that the gain is concentrated in output tokens per second rather than in time to first token. That distinction is worth internalising before you buy, because it decides which workloads benefit. A batch job that waits nine seconds for a first token and then streams for four minutes gains a great deal. An autocomplete path that needs a first token inside 200 milliseconds gains nothing at all, and pays double for the privilege. Anthropic's own framing points at the same conclusion — fast mode is for interactive work like rapid iteration and live debugging, and the gain is most visible under streaming, where there is more output to speed up.

There is a second-order effect that matters more than the headline for long sessions. Because the gain is on output, the wall-clock saving is proportional to how much the model writes, not to how long the conversation is. A model that spends its time thinking before it types is a model that fast mode cannot do much for, and Claude Opus 5.5 is a model that thinks by default. Adaptive thinking on Opus 5.5 is always on and cannot be disabled — a request that tries to set thinking to disabled returns a 400 error, and depth is controlled with the effort parameter instead. Fast mode makes the writing faster. It does not make the thinking faster, and no request parameter currently does.

The rate card: $8 and $40, flat, with multipliers on top

Fast mode pricing on Claude Opus 5.5 is $8 per million input tokens and $40 per million output tokens. Against the standard $4 and $20, that is a straight doubling — not a premium tier with a blended discount, not a per-request surcharge, and not something that stops being available past a context threshold.

• Standard Claude Opus 5.5 — $4.00 per million input, $20.00 per million output, $0.20 cache read, $5.00 cache write

• Fast mode Claude Opus 5.5 — $8.00 per million input, $40.00 per million output, flat across the full context window

• Batch API Claude Opus 5.5 — $2.00 per million input, $10.00 per million output, a 50 percent discount for asynchronous work

• Fast mode on Claude Opus 5 and Claude Opus 4.8 — $10.00 and $50.00 per million, still research preview

• Not available — fast mode cannot be combined with the Batch API at all, and cache-read pricing on the standard rate is currently the cheapest input path Anthropic sells on this model

Two modifiers stack on top of the fast rate rather than replacing it. Prompt-caching multipliers apply to fast mode pricing, and so do the data-residency multipliers that apply when a request pins an inference geography. Both are multipliers on a rate that is already twice standard, so a fast-mode request with a cache write and a residency multiplier is not a two-times-cost request. It is several times cost, and that is the shape of the bill nobody budgets for.

The flatness across the context window deserves its own sentence, because it is unusual. Fast mode at $8 and $40 applies to a 900,000-token request at the same rate as to a 9,000-token one. On a model with no long-context surcharge at standard speed, fast mode is therefore a clean 2x with no cliff — which makes the arithmetic easy and, unfortunately, makes it easy to spend without noticing.

The subscription question nobody answered clearly

The part of this change that is going to surprise the most people has nothing to do with the API. In Claude Code, fast mode is available to subscribers on Pro, Max, Team and Enterprise plans — and it is billed from usage credits, not from your plan's rate limits. The documentation is unambiguous: for Claude Code users on subscription plans, fast mode is available via usage credits only and is not included in the subscription rate limits.

So the practical answer to "does my Max plan cover this" is no. Your subscription covers standard-speed Claude Opus 5.5 inside your plan limits. Fast mode sits outside them and meters against credits you buy separately. Three further conditions apply:

• Team and Enterprise — an Owner has to enable fast mode for the organisation before anyone on the plan can use it

• Claude Console — the organisation needs fast mode access provisioned, and the spend then appears as ordinary API usage rather than credits

• Cloud sessions — fast mode is available for the session only and is not saved as your default; on claude.ai/code it appears as a switch in the model menu when your plan includes it

One thing worth remembering when you look at the credits balance: when Anthropic launched Claude Opus 5.5 in September it also raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and issued subscription users a rate-limit reset they can save and spend whenever they choose. That reset is the correct thing to spend on a fast-mode experiment. Burn the reset, measure the credits fast mode actually eats on your workload, and then decide whether the doubling is worth it — instead of discovering the answer on an invoice.

Turning it on, and the one cost trap worth knowing

In the Claude Code CLI, the toggle is /fast. Press space to switch it, press enter to confirm, and a small ↯ icon appears next to the prompt while it is active. There is a Toggle fast mode command in the VS Code extension, and a fastMode setting in your user settings file for anyone who wants it on permanently. The choice persists across sessions by default, and it can be configured to require a fresh opt-in each session instead.

On the API side it is a request parameter: send speed: "fast" and the response reports what you actually got in usage.speed, which reads "fast" or "standard". That field is worth logging, because it is the difference between a request that ran fast and a request that silently did not. Fast mode has its own rate limit, separate from the standard Opus limit, and exceeding it returns a 429 with a retry-after header rather than quietly falling back. A request for fast mode on a model that cannot serve it returns an error.

Then the trap. The first time you enable fast mode inside a conversation, you pay the full uncached fast-mode input price on the entire conversation context — every token already in the window. The deeper into a session you switch it on, the larger that one-time charge is. The cost applies once per conversation, so toggling off and back on later does not repeat it, but switching on at turn forty of a long agent run is materially more expensive than switching on at turn one. Anthropic's own guidance is to enable fast mode at the start of a session rather than mid-conversation, and for anyone running long agentic sessions that is not a style tip, it is the difference between a 2x bill and a much larger one.

Where a router fits, and where it does not

Fast mode is a request parameter on Anthropic's own API surface, and it is worth being straight about the consequence: the catalogue entry for Claude Opus 5.5 on OrcaRouter exposes the standard parameter set — reasoning effort, tools, structured outputs, verbosity, temperature — and fast mode is not among the parameters that entry advertises. The honest reading is that the 2x speed option is bought at Anthropic's door, and the place a routing layer earns its keep here is the decision around it.

That decision is usually the same one. Fast mode is a flat 2x on every token, and most production systems have a small latency-critical slice and a large everything-else slice. Paying 2x across the board to serve a slice is the expensive way to solve it. OrcaRouter routes Claude Opus 5.5 as anthropic/claude-opus-5.5 at Anthropic's published list price with 0% markup — the provider's rate passed through untouched, so the $4.00 input and $20.00 output figures above are what lands on the bill rather than a marked-up equivalent — and the routing DSL expresses the split as configuration: one rule sends the interactive slice to a cheaper, faster model, and the rest stays on Claude Opus 5.5 at standard speed. Automatic failover covers the other half of the problem, degrading a failed call to another provider path instead of to an error page. Across more than 200 models, one key, one bill, and no second contract to sign when the answer to "is this fast enough" turns out to be a different model rather than a faster setting.

Who should buy it this week, and who should not

Buy it if the bottleneck in your loop is the model writing, not the model thinking. Live debugging, rapid iteration on a prompt, an agent whose turns are dominated by emitting a long tool argument or a long code edit — those are the workloads where 2.5x output speed converts directly into a shorter wall clock, and where the doubled rate buys something a reader can feel.

Do not buy it to make a batch pipeline finish sooner; the Batch API is explicitly incompatible with fast mode and already sells the same model at half of standard price, which is a quarter of the fast rate. Do not buy it for classification, extraction or routing, where output is short, the caching economics are the dominant term, and the standard rate plus a cache read at $0.20 per million is doing the real work. And do not buy it because the queue opened. The queue opening is an access change, not a capability change, and a model that was fast enough for your workload last week at $4 and $20 has not become insufficient at $8 and $40.

What is genuinely new is that the trial cost has fallen to zero. There is no approval to wait for and no waitlist to sit on, so the correct response to October 15 is a measurement rather than a purchase: turn fast mode on at the start of a conversation, note what usage.speed reports, run your own interactive workload both ways, and let the credits decide. The one thing to watch next is the other two Opus models. Fast mode for Claude Opus 5 and Claude Opus 4.8 is still in research preview with the beta header still required, and if Anthropic extends the October 15 treatment to them, the premium-tier line-up gets simpler and the choice gets harder.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.