What Paying for a Faster AI Model Actually Costs You

The same model, served faster, is now a product with its own price. In the week to September 25, 2026 Xiaomi, Z.ai and Anthropic each put a faster version of an existing model on a price list, and Alibaba's Prime tiers surfaced on its pricing pages, joining the fast and priority tiers OpenAI, Google, xAI, Moonshot and Mistral already sell. The multiple is not standard. Most tiers cost exactly twice the base. Google's cost 1.8 times. One costs ten times. And we found no measurement of any of them by anyone other than the vendor.

This census puts every priced tier we could find on one table: the multiple, the vendor's speed claim in the vendor's own unit, and what else changes when you flip the flag, because a faster tier usually loses something too, such as batch pricing, a cache, a region or a rate limit. Then a decision rule for when a multiple is worth paying and when a smaller model is the cheaper answer to the same latency problem.

The scope rule for the table is strict: a row qualifies only when one vendor sells the same model at two service levels for two prices. A different model whose name contains "Fast" or "Turbo" is not a speed tier, it is a different model, and is excluded. Cheaper-and-slower tiers such as OpenAI Flex and Gemini Flex are listed below the table for contrast, because on one model they show the full spread: Gemini 3.8 Flash costs 0.5× on Flex and 1.8× on Priority, so the fastest synchronous tier costs 3.6 times the cheapest.

Multiples are the tier price divided by the base price at the same unit and context band, computed from each vendor's pricing page. Speed claims are quoted as the vendor wrote them. The fourth column is what else the vendor's documentation says changes.

Two contrast rows sit outside the census because they run the other way. OpenAI Flex serves GPT-6 Astra at $5 and $25, half of standard, with the documentation warning of slower responses and occasional unavailability that returns an uncharged error. Gemini Flex, in preview, is also 0.5× with a stated target of one to fifteen minutes and no automatic upgrade to standard. The batch tiers across vendors are a separate, mostly 0.5×, market that our batch pricing landscape covers.

The Xiaomi row is the one people will quote, and it needs its context. The base model is cheap: MiMo-V2.6-Pro is $0.435 input and $0.87 output, so ten times that is $4.35 and $8.70, in the same band as a frontier model's standard tier. Our post on the MiMo-V2.6 release covers what Xiaomi published about the model itself.

The rows split into two kinds of product, and vendors use the same word, priority, for both. The first kind is a faster serving configuration: Anthropic describes fast mode as the same model weights on a faster inference configuration, xAI calls Grok 4.7 Fast the same model on faster infrastructure, and Z.ai and Moonshot quote absolute token rates. The second kind is queue position: Gemini's priority inference, xAI's priority processing and Mistral's priority tier promise that your request is served ahead of standard traffic under load, which lowers time to first token when the queue is long and does nothing when it is empty. OpenAI's Fast mode describes both faster speeds and more consistent latency.

That distinction decides whether a benchmark on a quiet afternoon tells you anything. A serving-configuration tier should show its speedup at any hour. A queue-priority tier shows nothing until the vendor is busy, which is also when you cannot reproduce the test. Our latency benchmark reference explains why time to first token and output rate must be reported separately, and the table above shows that vendors already choose which one to promise.

Five other things change with the flag, and each is in a row above. Batch drops out at Anthropic, Xiaomi and xAI, and Alibaba's Prime rows lose the batch and caching tags their base rows carry. The cache can reset: Anthropic's fast and standard requests do not share cached prefixes, so a fallback to standard on a long agent session pays the cache write again. Rate limits move in both directions: Anthropic gives fast mode its own limit, Gemini Priority defaults to 0.3× standard, OpenAI shares the standard limit and downgrades traffic that ramps too quickly. Availability narrows, sometimes to a single region or a pair of clients. And the premium can silently not apply: OpenAI, Google, xAI and Mistral all serve over-limit or ramping traffic at standard and bill it standard, recorded only in a response field. Anthropic's Opus 4.6 goes further and accepts the fast flag while running at standard speed and price, with no error.

A speed tier is a way to buy latency without changing the answer. A smaller model is a way to buy latency by changing it. The right choice depends on which part of the latency you are paying for and whether the task can tolerate a weaker model at all.

The arithmetic that settles most of these is simple. If a tier costs 2× and the vendor claims up to 2.5× output speed, then at best you get 25% more output speed per dollar and at worst, when the claim is a ceiling, you pay more. A smaller model at a fifth of the price that is twice as fast wins on both axes whenever its answers are good enough. Our active-parameters reference has the architecture side of that comparison and our price index the standard-tier prices.

No vendor in the table publishes a measurement method for its speed claim: not the prompt length, the output length, the batch size, the region or the hour. Anthropic comes closest by naming the metric it targets and the one it does not. Google's documentation gives a latency band rather than a number. Xiaomi gives a multiple with no unit. Until one of them publishes a method, the right way to quote any of these figures is the way this table does: in quotation marks, with the vendor's name.

Not verified and therefore not in the census: Vertex AI's own provisioned-throughput pricing; a third-party host's fast endpoint for GLM-5.3 that appears on an aggregator at 1.5× but has no vendor documentation we could read; any DeepSeek speed tier, because none exists on its pricing page; Kimi K3, which has no high-speed SKU; and any fast mode for Claude Fable 5.1 or Sonnet 5, which Anthropic's fast-mode page does not list. Alibaba's English and Chinese Prime pages disagree on which models are included, and the English Prime page and the pricing page disagree on one regional video price, so the Prime rows cite the page each figure came from.