About TierCliff
A one-person project that prices a single OpenAI request the way the invoice will: tier, caching, reasoning effort and the 272K repricing threshold together.
The name comes from a threshold. On the current GPT-6 models, a request whose input crosses 272,000 tokens is not charged a little more for the extra tokens. The whole request is re-priced: input and cached input double, and output rises by half. The bill does not slope upward there, it steps off a ledge, and an agent that appends its history every turn can walk off that ledge without anyone deciding to.
That threshold is easy to find in OpenAI's documentation. What is hard to find is a single place that combines it with everything else that moves the price of one request: the service tier, the cache hit rate, the cache-write charge, the reasoning effort and the regional-processing surcharge. TierCliff is that place. You describe one request and how often it runs, and it gives you the per-call and monthly figure with every multiplier printed next to it.
The four things the calculator multiplies together
| lever | what it does to the rate | published or assumed |
|---|---|---|
| Service tier | Batch and Flex 0.5x, Standard 1x, Fast 2x, Ultrafast 6x | published |
| Long context | above 272K input, input and cache 2x, output 1.5x, on the whole request | published |
| Prompt caching | cached input at 5% to 10% of the input rate depending on the model, plus a write charge | published |
| Reasoning effort | how many hidden reasoning tokens bill as output | assumed, editable |
The first three come from OpenAI's pricing documentation and model pages. The fourth is ours. OpenAI does not publish how many reasoning tokens each effort level produces, so the calculator uses planning multipliers, labels them as assumptions everywhere they appear, and lets you replace them with figures from your own usage reports.
What it does not try to do
It does not benchmark models, rank them on quality or tell you which one is smartest. It does not model rate limits, retries or the cost of a failed request. It does not cover embeddings, images, audio or fine-tuning. Those are real costs, but each has its own pricing shape, and adding them would turn a precise tool for one question into a vague tool for many.
Who runs it
TierCliff is built and maintained by Charles, one developer, in spare time. There is no team behind it and no company. It is not affiliated with, endorsed by or sponsored by OpenAI. The calculator runs entirely in your browser: there are no accounts, and nothing you type is sent to a server.
The site is paid for by display advertising. There is no sponsored content, no affiliate link to OpenAI and no paid placement, and advertising does not change which tier the tool suggests for a request.
Corrections log
Pricing pages go wrong quietly, so every substantive correction is listed here with its date rather than silently overwritten.
- 27 October 2026. GPT-6.1 Sol Ultrafast became available to all API users on 8 October; the calculator and the Ultrafast guide had still described it as announced but not open. Luna, GPT-5.6 Sol and GPT-5.6 Terra accept the full range of effort levels, not only low to high; the calculator now offers all of them. A maximum-input figure that could not be traced to an official page was removed. The three "open questions" previously listed on the methodology page were each settled by OpenAI's own documentation and the answers are now published there.
- 2 October 2026. The site grew from a single calculator into a reference: tier, caching and long-context guides, a methodology page and this page.
If you find something wrong, the contact page explains what to send. Corrections are the most useful message this site gets.
Where to start
- The calculator for a figure on your own workload
- The 272K threshold for the repricing rule the site is named after
- Service tiers for what each tier costs and when it is the wrong choice
- Methodology for every rate, its source and the formula