Methodology — where every number on this site comes from
The formula the calculator runs, the official pages each rate comes from, and a plain list of what is assumed rather than published.
The calculator turns one request description into a dollar figure. This page shows how, so you can check any number on the site against OpenAI's own documentation, and see exactly where the site is making an assumption rather than quoting a rate.
How one request is priced
1 base = uncached_input x input_rate
+ cache_read x cache_read_rate
+ cache_write x cache_write_rate
+ output_tokens x output_rate
output_tokens = visible output x effort multiplier
2 long if input > 272,000 tokens, every line above is multiplied
by the model's long-context factor, for the WHOLE request:
input x2, cache read x2, cache write x2, output x1.5
3 tier multiply by the service tier:
batch 0.5, flex 0.5, standard 1, fast 2, ultrafast 6
4 region multiply by 1.10 if the request runs on a regional endpoint
Step 2 is the one most estimates get wrong. The long-context factor is not charged on the tokens above 272,000; it re-prices the entire request. A request at 273,000 input tokens is billed as though all 273,000 were long-context tokens. That discontinuity is what the site is named after.
Where the published figures come from
Every rate is read from OpenAI's own pages. No third-party price table is used as a source.
| page | what it provides |
|---|---|
| API pricing | base rates, cache rates, tier tables, the regional uplift |
| Model pages | long-context multipliers, context window, supported effort levels, regional notes |
| Fast mode guide | the 2x rate, the up-to-2.5x speed figure, EU residency per model |
| Ultrafast mode guide | which models offer it, regions, separate rate limits |
| Flex processing guide | Flex priced at Batch rates, cache discounts included |
| Changelog | release and availability dates |
All of it was last checked on 27 October 2026. The rates live in one data file that every page reads, and the footer of every page prints the verification date, so a stale figure is visible rather than silent.
What is assumed, and which way each assumption leans
These are the places where OpenAI's documentation does not give a number, and the calculator has to pick one. Each is listed with the direction it errs in.
| assumption | what the calculator does | why |
|---|---|---|
| Reasoning tokens per effort level | low x1, medium x3, high x8, xhigh x20, max x45 of visible output | OpenAI publishes no ratio. Editable in the calculator; replace with ratios from your own usage reports. |
| Ultrafast above 272K input | the 6x tier and the long-context factor are multiplied together | not stated either way; multiplying gives the higher, safer estimate |
| Regional uplift on a non-Standard tier | the 1.10 is multiplied with the tier rate | the uplift is stated per request without naming tiers |
| Long-context factor on cache lines for GPT-5.6 models | cache read and write are doubled, like input | the GPT-5.6 pages state "2x input, 1.5x output" only; GPT-6 pages state the cache lines explicitly |
| Cache-write lifetime | the standard short-lifetime write rate | the 12-hour cache (GPT-5.6 and later, Responses API only) is billed at a higher write rate and is not modelled |
Where availability differs by model
Tier availability on a regional endpoint is set per model, not per tier. The calculator disables a tier on the regional toggle only where OpenAI says it is unavailable, or where nothing is published for that model.
| model | Fast on EU endpoint | Ultrafast | Batch and Flex on EU endpoint |
|---|---|---|---|
| GPT-6.1 Sol | yes | yes, including EU | not stated; allowed in the calculator |
| GPT-6 Astra | no | yes, US and global only | not stated; allowed in the calculator |
| GPT-6 Luna | yes | not offered | yes |
| GPT-5.6 Sol, GPT-5.6 Terra | not stated; disabled in the calculator | not offered | not stated; allowed in the calculator |
All five models accept the effort levels low through max. Luna and the GPT-5.6 models also accept a none setting, which the calculator does not list because it produces no reasoning tokens to price.
Questions this page used to leave open
Earlier versions listed three pricing interactions as unconfirmed. OpenAI's documentation now settles all three:
- Do Batch and Flex discounts apply to cached input? Yes. Flex is priced at Batch rates with prompt-caching discounts on top, and the Batch/Flex table shows cached rates at half of Standard.
- Does the longer-lifetime cache write cost more on every model? On GPT-5.6 and later, the 30-minute cache write is 1.25x input and the 12-hour write is 2.5x, with the 12-hour option only in the Responses API. Earlier models use a different retention setting.
- Which rate limits does Ultrafast use? Its own. Ultrafast has rate limits separate from Standard and Fast, while Fast draws from the same limits as Standard.
What the calculator leaves out
- Retries and failures. A failed request that is re-sent costs twice. Only successful requests are counted.
- Rate limits and queueing. Flex may queue or be unavailable; that is described, not priced.
- Latency. The faster tiers are priced, but how much time they save is yours to measure. OpenAI publishes a speed ceiling for Fast and none for Ultrafast.
- Everything that is not a text request: embeddings, images, audio, fine-tuning and storage.
Corrections and independence
When a figure here turns out to be wrong, the page is corrected and the change is logged with its date on the about page. Corrections are welcome through the contact page.
TierCliff is not affiliated with, endorsed by or sponsored by OpenAI. Product names and prices belong to their respective owners. The site is funded by advertising, and no advertiser has any influence on what the calculator recommends.