Research preview — the rate can move without notice
The Fast mode rates on this page were read from Anthropic's own pricing page and Fast mode documentation on undefined. The standard base rates our registry holds for these models were last checked on undefined. Anthropic labels Fast mode a gated research preview, so treat the table below as current, not contractual.
Claude Opus 5 specs and every API rateResearch preview · Claude Opus 5 & Opus 4.8
Claude Fast Mode Pricing
Fast mode trades money for latency on Claude Opus 5 and Claude Opus 4.8. Here is Anthropic's published premium and speed claim — up to 2.5x higher output tokens per second — how it stacks with prompt caching, long context and the Batch API, and the arithmetic that settles whether it is worth it.
The short answer
Anthropic bills Claude Fast mode at $10 per million input tokens and $50 per million output tokens for Claude Opus 5 and Claude Opus 4.8. It is a gated research preview on the Claude API, including Claude Managed Agents, and nowhere else: you request access through your account manager, or join the waitlist if you do not have one. What you get is published — up to 2.5x higher output tokens per second, aimed at output speed rather than time to first token, with no change to intelligence or capabilities. The Fast rate holds across the full context window, including requests over 200K input tokens, and prompt-caching and data-residency multipliers apply on top of it.
What is Claude Fast mode?
Fast mode is a pricing option you turn on per request, not a different model: set speed: "fast" together with the fast-mode-2026-02-01 beta header on the call you already make. The model stays the same; only the rate and the inference configuration change.
A gated research preview, priced per token
Anthropic's pricing page describes Fast mode as providing significantly faster output for Claude Opus 5 and Claude Opus 4.8 at premium pricing, and its Fast mode doc puts the feature in research preview with access on request: ask your account manager, or join the waitlist if you do not have one. Billing is a separate rate line, so it changes what a million tokens costs you rather than which model answers.
Up to 2.5x more output tokens per second
Anthropic's Fast mode doc states the mode delivers up to 2.5x higher output tokens per second from Claude Opus 5 and Claude Opus 4.8, and that the benefit is focused on output tokens per second (OTPS), not time to first token (TTFT) — a long prompt still takes as long to start. The same doc is explicit that nothing else moves: Fast mode runs the same model with a faster inference configuration, with no change to intelligence or capabilities. Read 2.5x as a ceiling, not a service level: no latency percentile and no typical-case figure is published, so measure both modes on your own workload before you move traffic.
Who it is for
Interactive work where a person is waiting on the response: chat interfaces, IDE and agent loops, autocomplete, extraction inside a request path. Background work is the wrong fit — that is what the Batch API is for, and Anthropic states Fast mode is not available with Batch. And since access is granted per account during the preview, confirm you can get it before designing around it.
Claude Opus 5 specs and API ratesClaude API: endpoints, auth, request shape
Claude Fast mode price per million tokens
Standard rates come from our model registry; the Fast mode column is Anthropic's published rate for the preview. The last column is our own division of the two, not vendor wording.
| Model | Standard rate (in / out, per MTok) | Fast mode rate (in / out, per MTok) | Fast vs standard (in / out) |
|---|---|---|---|
| Claude Opus 5 | $5.00 / $25.00 | $10.00 / $50.00 | 2x / 2x |
| Claude Opus 4.8 | $5.00 / $25.00 | $10.00 / $50.00 | 2x / 2x |
| Claude Opus 4.7 | Not in our registry | Returns an error, with no fallback to standard speed | — |
| Claude Opus 4.6 | Not in our registry | Standard speed, standard rates, usage.speed: "standard" | — |
One Fast mode line covers both supported models; the multiple against each model's own standard rate sits in the last column, computed from the registry. Opus 4.7 behaves differently: a request carrying the speed: "fast" parameter returns an error and, unlike Opus 4.6, does not fall back to standard speed — the model itself stays available at standard speed. Opus 4.6 does not error at all: those requests run at standard speed, are billed at standard rates, and the response reports usage.speed: "standard". For both, Anthropic's instruction is to migrate to Claude Opus 5 or Claude Opus 4.8 to keep using Fast mode. The two older standard prices stay unquoted here: Anthropic publishes $5 input / $25 output per million tokens for each on the same pricing page, but our model registry holds no row for Opus 4.7 or Opus 4.6, and this table prints only figures it can re-derive from the registry.
How Fast mode stacks with caching, context and Batch
Fast mode replaces the per-token rate; the other modifiers keep working on top of it. Anthropic publishes the rules — the combined figures below are our arithmetic.
Prompt caching discounts the Fast rate
Cache rates are published as multipliers of the input price, and Anthropic states the prompt-caching multipliers apply on top of Fast mode pricing. So a cache read is discounted relative to the Fast input rate, not the standard one — and cache writes follow the same rule.
Anthropic publishes no long-context premium
Two lines on Anthropic's pricing page decide this row. Fast mode pricing applies across the full context window, including requests over 200k input tokens, so the Fast rate does not climb as your prompt gets long. And Claude 4.6 and later models include the full context window at standard pricing — a 900k-token request is billed at the same per-token rate as a 9k-token one, with caching and batch discounts applying across the whole window. That is a real difference at this tier: the rival rows in our registry record xAI doubling both Grok 4.7 rates at prompts of 200K tokens and above, and OpenAI and Sakana billing 2x input / 1.5x output above 272K on GPT-6 Astra and Fugu Ultra v2. Anthropic says nothing about a size tier for models older than 4.6, so this claim stops at that line.
Not available with the Batch API
The two are mutually exclusive. Batch bills below the standard rate and is slow by design; Fast mode bills above it and is quick. If a job can wait, the discount is on the other side of this choice.
Data-residency multipliers apply on top
Anthropic publishes the size this time: regional and multi-region endpoints carry a 10% premium over global endpoints, and the first-party Claude API is global by default. The same page states that data-residency multipliers apply on top of Fast mode pricing, so the premium lands on the Fast rate rather than the standard one. Anthropic scopes the residency structure to Claude Sonnet 4.5, Haiku 4.5, Opus 4.5 and all later models, which covers both Fast mode models.
| Modifier, applied to Claude Opus 5 | Standard rate (registry) | Under Fast mode (derived) |
|---|---|---|
| Cache read / refresh | $0.50 | $1.00 |
| 5-minute cache write | $6.25 | $12.50 |
| 1-hour cache write | $10.00 | $20.00 |
| Batch API (in / out) | $2.50 / $12.50 | Not available with Fast mode |
| Requests over 200K input tokens | $5.00 / $25.00 across the full window | $10.00 / $50.00 across the full window |
| Data-residency multiplier | +10% over a global endpoint | Multiplier applies on top |
The Fast mode column is derived, not published: we read each modifier's multiplier off the Opus 5 registry row (cache read divided by base input, cache write divided by base input — the 1-hour write included now that the registry carries it) and apply it to the Fast input rate, because Anthropic says caching and residency multipliers stack on top of Fast mode pricing. Anthropic prints no such combined figure, so treat these as budget estimates and check your invoice. The over-200K row is the exception: both rates there are published and neither climbs with prompt length, and the residency cell shows Anthropic's published 10% endpoint premium rather than a computed price.
Where Claude Fast mode is available
Four things decide whether you can actually use it: the platforms Anthropic lists, the ones it rules out, the ones it stays silent about, and how access to the preview is granted.
The Claude API, including Claude Managed Agents, only
Anthropic states the scope plainly: Fast mode is available as a research preview on the Claude API, including Claude Managed Agents, only. Managed Agents count as the same first-party surface, so the Fast rate is set and served by Anthropic's own API — anything reached through another provider's console is outside it.
Not on Bedrock, AWS, Google Cloud or Foundry
The Fast mode doc names the exclusions: it is not available on Amazon Bedrock, Claude Platform on AWS, Google Cloud, or Microsoft Foundry. The pricing page says the same thing in other words — first-party Claude API only, no partner-operated platforms. If your contract reaches Claude through one of those four, plan on standard speed and standard pricing there; Anthropic lists them as places Fast mode is not available, not as a later rollout.
The Claude apps are not on the availability list
Anthropic's published list covers the Claude API and Claude Managed Agents and does not mention the Claude apps at all — neither the free tier nor Pro or Max. That is an absence of documentation, not a confirmed refusal, so this page records the apps as outside the stated scope rather than ruling them out. What also stays unconfirmed is whether an app-level toggle exists and how it would draw on plan limits: no page we verified describes one.
A gated research preview
Access is not self-serve: Anthropic asks you to contact your account manager to request Fast mode, and to join the waitlist if you do not have an account manager. The preview has no published end date, and while it runs the rate, the supported model list and the platform availability can all change. Re-read the vendor's pricing and Fast mode docs before you build a budget on these numbers.
What the Fast mode rate buys in the lineup
The premium looks different depending on which model you compare it against. All rates below are registry figures per million tokens.
| Model and mode | Input / MTok | Output / MTok |
|---|---|---|
| Claude Opus 5 — Fast mode | $10.00 | $50.00 |
| Claude Fable 5.1 — standard rate | $10.00 | $50.00 |
| Claude Opus 5 — standard rate | $5.00 | $25.00 |
| Claude Sonnet 5 — standard rate | $2.00 | $10.00 |
| Claude Haiku 4.5 — standard rate | $1.00 | $5.00 |
Priced per token, Fast mode on Opus 5 lands on the same input and output rates as Claude Fable 5.1 at standard pricing: same money per million tokens, different model. That is the comparison worth making before you buy speed — whether the model you would otherwise run at standard rates is the better answer for the same bill.
Is Claude Fast mode worth the premium?
It is a straight multiple on the token bill, so the question has an arithmetic answer: what is the saved wall-clock time worth per request?
Modelled workload: 500 requests a month, 10K input and 1K output tokens each, priced on Claude Opus 5 with no cache hits.
- Standard rate
- $37.50
- Fast mode rate
- $75.00
- Extra per month
- +$37.50
- Extra per request
- +$0.075
The decision rule
Pay the premium when a person is waiting on the response and the extra cost of one more request is small next to the time it saves. Skip it when nobody is watching the clock: batch the work at the discounted rate, shorten the prompt, or move the load-bearing part of the traffic to a cheaper model and keep Fast mode for the steps users actually feel.
The example bills every input token fresh, so a cached prefix changes the totals — caching discounts the Fast rate rather than removing it. The speed side is still a ceiling rather than a number you can plan against: Anthropic publishes up to 2.5x higher output tokens per second, no latency percentile and no typical case, and none of it is a promise about your request. So this page prices the premium exactly and tells you what changes on the time side only in direction.
Model your own monthly token billClaude model specs side by side
Every figure here is either read from Anthropic's own pricing and Fast mode docs on 2026-09-22 or computed from the model registry that backs this site, whose Anthropic rates were last checked on 2026-09-22. What is still open after this pass: whether the Claude apps expose Fast mode at all, the pricing terms of zero-data-retention for it, when the preview ends, whether Claude models older than 4.6 carry a long-context tier, and any latency percentile behind the 2.5x output-speed claim. Those cells say they are unconfirmed instead of guessing.
Price the speed before you buy it
Put your own token counts next to the rates above and see what the premium does to a month of traffic, or read the full pricing table across every model we track.