The same model from different sellers: what to check
When several sellers offer one model, the lowest token rate is a start; context window, serving tier, cached and batch rates and limits decide the real cost.
Check more than the rate when one model has several sellers
When the same model is sold by several providers, compare the input and output rates first, then check the context window each seller allows, whether a cached or batch rate exists, which serving tier the row is, and the rate limits on your account. Sellers can list one model name and still sell different products. The rate tells you which row to read first, not which to pick.
Models priced here: 467. Models among them sold by more than one seller: 93. Those are the ones where choosing a seller is a real choice, and the token price table shows each seller's rates side by side.
A model is matched across sellers by its name and version. Where sellers spell a model differently or ship a variant under a new name, it can appear as separate models until the names are reconciled, so a model with one seller listed may still be sold elsewhere.
Why one model name has several prices
Sellers run models on their own hardware, with their own serving software and their own margins. A seller may serve a model at a lower precision, cap the context window below what the model supports, or bundle a higher rate limit into the price.
None of that is visible in a per-million rate, and some of it is not published at all. A gap between sellers is a prompt to read both pages, not proof that the dearer one is overcharging or that the cheaper one is cutting corners.
Rates also move. A seller that held a low rate last month may have changed it, and the row will show the new figure once the page is read again. The date beside each rate says how fresh it is.
A checklist for each row
Context window: does this seller accept the longest prompt you send? A seller with a lower rate and a shorter window may not run your workload at all. Precision: does the seller say whether the weights are served in full or reduced precision? Where the page does not say, ask, and test your own prompts on both before committing.
Tier: is this row the standard interactive tier, a batch tier or a free tier? A free row with tight limits is not the same product as a paid row. Cached and batch rates: does this seller publish them for this model? A seller with a slightly higher standard rate and a cached rate can be cheaper for a workload with a long repeated prefix.
Limits and terms: requests and tokens per minute on a new account, data retention, and where requests are processed. Those are set per seller and per account, and this site does not record them.
Running the comparison on your own mix
For each seller, monthly cost is input tokens over a million times that seller's input rate, plus output tokens over a million times its output rate, adjusted for the cached share where a cached rate exists. Do it at your own token totals, because a ranking on input rate alone can flip once output is counted.
The LLM cost calculator lays this out as one row per model and seller, cheapest first, with every other row showing how much more it costs. Pick the model, set the input, output and cached sliders to your measured month, and filter by serving tier to keep batch and free rows out of a comparison of interactive ones.
Then test. Send a sample of real prompts to the cheapest few rows and compare the outputs yourself. Differences that matter to you belong in the decision beside the rate.
Keeping the option to switch
Many sellers of a hosted model expose a compatible API, so moving between them can be a configuration change. A gateway or router in front of your calls makes that easier and lets you send traffic to a second seller when the first is rate limiting you.
Rates change. This site hides any rate older than 21 days rather than show a stale one, so a seller can drop out of the table until it is read again. Check the date beside a rate before quoting it to anyone.
Keep the comparison current as well. A quarterly rerun of the calculator at your latest token totals, with the same filters, shows whether a change of seller has become worth the effort since the last time anyone looked.
What the price spread says and does not say
A wide spread between sellers of one model is a prompt to look, not a verdict. It can reflect a shorter context window, a different precision, a higher rate limit included in the price, or simply a seller that has not repriced recently. The rate alone cannot say which.
A narrow spread is not proof that sellers are interchangeable either. Two rows at similar rates can differ on limits, retention and where requests run, all of which matter to some buyers more than the price does.
The spread is most useful as a ceiling on what switching could save. Close rates at the cheapest and dearest sellers of your model make the effort of moving hard to justify on price. Far-apart rates justify an afternoon on the comparison above.
What we could not answer
- ⚑We do not record serving precision, rate limits or data retention per seller, so the checklist names questions this site cannot answer for you.
- ⚑Nothing here compares the output of one model across sellers.
What this rests on
infratoapp, “The same model from different sellers: what to check”, read 27 October 2026. models sold by more than one seller: 93. https://www.infratoapp.com/blog/the-same-model-from-different-sellers
Questions people ask about this
Why does the same LLM cost different amounts at different providers?
Each seller runs the model on its own hardware and software, may serve it at a different precision or with a shorter context window, sets its own rate limits and chooses its own margin. The per-million rate reflects all of that, and most of it is not visible in the rate.
Is the cheapest provider for a model always the right one?
Not always. Check that the seller accepts your longest prompt, whether it publishes cached or batch rates that change your real cost, which tier the listing is, and its rate limits. Then send the same sample prompts to the cheapest few sellers and compare the outputs before committing.
Can I switch sellers for the same model easily?
Often, because many sellers expose a compatible API for hosted models, so a change can be a new endpoint and key. A gateway in front of your calls makes it simpler. Test outputs after switching, since precision and serving settings can differ between sellers of one model.
These render as FAQPage structured data · answers are filled from the same live figures
Read next
Token price data cannot tell open-weight models from closed ones here; it can compare what a workload costs per model and seller, which is the useful question.
A batch tier prices tokens lower in exchange for a delayed answer, so it fits work nobody is waiting on and fails anything a person or a pipeline is blocked by.
Input and output tokens carry separate per-million rates, output is commonly dearer, and a blended rate only means something for one fixed mix of the two.