Fourteen companies sell tokens from the same handful of models. The same weights are quoted at different prices by different hosts, so the choice is worth an hour.
Every price we track, in one table.
Pick the model first
Most providers in the table serve open weights: DeepSeek, Qwen, Kimi, GLM, MiMo. If your work runs on one of those, the model and the host are separate decisions. You can move between hosts without touching a prompt.
The closed models — OpenAI, Anthropic, Google, xAI, Mistral — come from one company each. There is no shopping around, only a choice of tier.
Then the bill, which is not the sticker price
A rate per million tokens is the start of the arithmetic. One task costs input tokens times the input rate, plus output tokens times the output rate, plus every retry.
Two things move that number more than the rate does:
- Cached input. Several providers charge a fraction of the input rate for a prefix they have already seen. An agent that re-sends the same system prompt a hundred times pays for it a hundred times unless caching applies.
- Reasoning tokens. On models that think before answering, the thinking bills as output. A model with a low output rate and long reasoning can cost more per answer than a model with a high rate and none.
The table's two price columns compare the rate. Estimating the two token counts for your own task takes an afternoon with real inputs, and it is the only number that decides anything.
Region and retention remove more rows than price
Region is where the provider runs inference, not where you are. Three regions appear in the table: the United States, the European Union and China.
If your data cannot leave a jurisdiction, that column cuts the list before price matters. Retention cuts it again. A provider that keeps prompts for abuse monitoring is not usable for some work at any price.
What to check before you sign
- The rate for your model, not the headline rate for the cheapest one.
- Whether the rate is introductory. Two rows in the table step up on a published date.
- Whether peak and off-peak rates apply, and which hours you actually run.
- The context limit, and whether the price changes above a threshold inside it.
- The date in the last column. A price older than six weeks is a guess.
Run two providers, not one
One provider is one point of failure: rate limits, an outage, a deprecation with two weeks' notice. A second provider for the same open model costs one extra key and one config line.
The table exists because the alternative is a spreadsheet that goes stale in a month. It is built from a file that records where each price came from and the day it was read. The table is the only place on this site a price is written down.
Questions
- Is the cheapest provider always right?
- No. The lowest rate in the table belongs to a small open model at a host whose queue may not suit you.
- Do I have to switch providers to switch models?
- Only for closed models. For open weights the model and the host are separate choices.
- How often do these prices change?
- Often enough that every row carries a date. Read the date before you trust the number.