Faster responses and regional processing cost more on public AI price lists. As of 19 October 2026, Anthropic lists a 1.1 times multiplier for US only inference on newer models, OpenAI lists a 10 percent uplift for regional processing, and all three list a faster tier at a higher rate.
Data residency multipliers
Anthropic's pricing page says that for Claude 4.6 and later models, specifying US only inference with the inference_geo parameter applies a 1.1 times multiplier to all token categories: input, output, cache writes and cache reads. Global routing, the default, uses standard pricing. The page says the multiplier also applies to the higher prices for prompts over 100,000 tokens on Haiku 5.5.
OpenAI's pricing page notes that regional processing endpoints, and its FedRAMP endpoints, carry a 10 percent uplift on eligible models.
If you have a legal reason to keep processing in one region, price it in. If you do not, the default is cheaper. Whether you need a region is a question for your counsel and your customers' contracts, not a pricing decision.
Faster tiers
- OpenAI lists a Fast column, which the page says was formerly called Priority. For gpt-6.1-sol it shows $4.00 input and $20.00 output per million tokens against $2.00 and $10.00 standard, which is double. For gpt-5.5 the Fast input is $12.50 against $5.00 standard.
- Google lists Priority pricing as higher than standard. For Gemini 3.8 Flash it shows $1.35 input and $6.75 output against $0.75 and $3.75.
- Anthropic documents a fast mode with its own pricing section and dedicated rate limits for some Opus models. Read that section for current figures, since this post did not capture them.
Ultrafast tiers also appear on OpenAI's page for two models at several times the standard rate.
What each premium buys
A faster tier buys lower latency, and in some cases more predictable capacity under load. It does not make the model smarter. Use it where a person is waiting and delay hurts, such as an interactive assistant. Do not use it for background jobs, which belong in batch at roughly half price. See batch API for LLM workloads.
Stacking multipliers
Anthropic says the data residency multiplier stacks with caching and batch modifiers. A job can therefore combine a regional multiplier, a cache discount and a batch discount. Work out the combined rate for one real request before you build a forecast. The worked example shows the arithmetic.
A budgeting rule
Put each premium behind a named reason. If you cannot say why a request needs a fast or regional path, send it down the cheaper one. Review the list when prices change. See scheduled price changes in AI APIs.
Quick check
Is regional processing a compliance guarantee?
No. A price page describes where inference runs. What your obligations are, and whether a setting meets them, needs a separate review.
Is fast mode the same as a cheaper small model?
No. Fast tiers run the same model sooner and cost more. A smaller model is a different model at a lower price.
Sources
Last reviewed: 19 October 2026.
