A token is a unit of billing, not a unit of value. What an inference provider sells is a place in a queue, a promise about where your data goes, and a rate that holds until the provider decides otherwise.
What each provider charges, and the day it was checked.
The queue is the product
Two companies running the same open weights produce the same tokens. What differs is how long you wait for the first one, and how many requests a minute you get before the door closes.
Cerebras and Groq sell silicon: their rows are open models at a rate that buys speed. DeepInfra and Together sell capacity: more models, lower rates, longer queues. Fireworks sells one model three ways — standard, spot and batch — and the price tracks how long you are willing to wait.
Reliability is a line item nobody quotes
Every provider in the table publishes a price per token. Few publish the number you actually need: the share of requests that finish inside your timeout.
An outage at three in the morning does not show up in a per-million rate. It shows up in retry logic, a second provider, and a page at three in the morning.
Region and retention
Region says where the model runs. For a European customer the EU row is not a preference, it is the contract. For a Chinese open model the cn row is the only route to the lab's own pricing.
Retention says what happens to a prompt after the answer. Some providers keep requests for abuse monitoring for a set number of days. Some offer a zero-retention arrangement for approved accounts. Neither is the default, and neither is free to arrange.
The right to change the price
Every provider reserves it. Two rows in the table step up on a published date. One charges a peak rate and half of it off-peak. One discount is running on the day we checked and will not run forever.
A budget built on today's row is a budget for today. That is why every line in the table carries a date, and why the job that rebuilds it fails when a row turns 45 days old.
Support is what the enterprise tier buys
The rate per token is the same at every tier. What changes above the self-serve plan is a person to email, an uptime commitment, and sometimes a data processing agreement.
What to ask before you commit
- Which rate applies to the model you actually call?
- Is the rate introductory, seasonal or spot?
- Where does inference run, and what does the provider keep?
- What happens at the rate limit?
- How much notice does a deprecation get?
The table answers the first two. The rest come from a week of your own traffic.
Questions
- Is a lower price per token always a lower bill?
- No. Retries, reasoning tokens and cache misses are billed too.
- Does region change the price?
- Sometimes. Each row is priced for the region it names.
- What does zero retention cost?
- Time and paperwork, usually, rather than a higher rate.