Twenty-nine prices from fourteen providers. One model, DeepSeek V4.1 Flash 1017, is sold by five of them at five different rates. Same weights, same million-token window, different bill.
The table: every row sourced, dated and linked to the page it came from.
The spread is the product
The cheapest input rate for that model in the table is two thirds of the dearest. Nobody is lying. Each company is pricing its own queue: how many GPUs it has, how full they are, which region they sit in, and what it thinks a customer will pay for the difference.
You are not buying the model. The weights are free to download. You are buying the host.
Model Brief runs on one of these
This site runs on DeepInfra, on DeepSeek V4.1 Flash 1017. We chose it on price and stayed for the context window. The rate we pay is the row in the table, and we re-read that page when we write these pieces.
No provider pays us. If one ever does, this paragraph changes before anything else does.
Where the money goes
The dearest output rate in the table is nearly two hundred times the cheapest. That ratio is not a frontier-versus-small-model story: both ends are real, shipping models, and both are on the same page.
Three things produce the spread:
- The model. A frontier closed model costs more per token than an open one, and that is most of the distance between the top and the bottom of the table.
- The service level. Fireworks sells the same open model three ways — standard, spot and batch — and the price follows how long you are willing to wait.
- The clock. DeepSeek charges one rate at peak and half of it off-peak. Google carries two rows that step up to a higher rate on a published date.
A router charges a fee and says so
OpenRouter resells models at their list price and takes its margin on the credits you buy: 5.5 percent on the Standard plan, 8 on Business. Astra appears twice in the table, once from OpenAI and once from OpenRouter, at the same rate. The router does not discount the model; it charges for one key, every model, and failover between hosts.
That fee is the honest shape of the deal. Compare it against the cost of keeping four accounts, four keys and four billing relationships current.
What a cheaper row does not buy
Throughput. Uptime. A support address. A rate that survives the quarter.
The table has seven columns and none of them are those. A row tells you what a token costs at a moment, from a page you can read yourself. Everything else about a provider comes from running your own work on it for a week.
Questions
- Which provider is cheapest for DeepSeek V4.1 Flash?
- DeepInfra at its list price, with a discount running on the day we checked. The number is in the table, with the date.
- Is OpenRouter more expensive?
- It charges the model's rate plus a fee on credits. What you get for the fee is one key for every model.
- Why does every row carry a date?
- Because a price without a date is a memory, and a memory is how budgets go wrong.