Open or closed models: what price data can and cannot say
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.
The price data cannot sort models into open and closed
The token prices on this site cannot tell you which models have open weights and which are closed, because the data does not record it reliably. What the data can do is price a workload on each model and seller from published rates. For a cost decision that is the question worth asking: what does this workload cost here, and what else do I take on to run it this way.
That limit is stated plainly because the open versus closed comparison is one buyers search for, and an answer built on a field that cannot say closed would look certain and be wrong. The token price table lists models by rate (models priced: 467), with no column for licence.
Licences are also more varied than the two labels suggest. Some weights can be downloaded but carry limits on commercial use; some models are hosted only by their maker; some are both. A single open or closed flag would hide most of that, even if the data held one.
What a per-token rate already contains
A hosted model sold per token, open weights or not, bundles the hardware, the serving software, the people running it and the seller's margin into one rate. When you compare two per-token rates you are comparing two bundles, and the licence of the weights is one ingredient among many.
Where a model is sold by several sellers, the rates between them show how much the bundle varies for one set of weights. Models with more than one seller on record: 93. On those, the data can make the comparison honestly.
For a model sold only by its maker, there is one bundle and no spread. For a model sold by many, the lowest rate is a reference point for anyone pricing a self-hosted plan of the same weights.
Where weights matter to cost
Open weights change cost when you run the model yourself. Then the bill is no longer tokens times a rate but GPU hours times an hourly rate, and the token count only matters through how many cards it takes to serve it. The GPU cost calculator prices that side from published hourly rates.
The comparison works out in words. Self-hosted monthly cost is cards needed times hours running times the hourly rate per card, plus the engineering time to operate it. API monthly cost is tokens in and out times their rates. Self-hosting tends to pay only when the cards stay busy, because an idle card is billed by the hour and an idle API costs nothing.
Memory sets the card count. A model's weights need about its parameter count times the bytes per parameter: two bytes in 16-bit, one in 8-bit, half in 4-bit. The key-value cache and activations need more on top, scale with context and batch size, and cannot be given as one figure.
Questions to answer outside the price table
Licence: read the model's own licence for what it allows, including any limits on commercial use or on the size of the company using it. Data: ask each seller what it keeps and where it processes requests. Control: decide whether you must pin an exact model version or fine-tune it, which may require weights you can hold.
None of these is in the price data, and several matter more than the rate for some buyers. A regulated workload may rule out every hosted option before cost is discussed.
Support is another. A hosted model comes with a seller who answers when it breaks. Self-hosted weights come with whoever on your team is on call. That person's time is a cost and belongs in the comparison.
A cost frame that survives the licence question
Write the workload down first: requests per month, input and output tokens per request, cached share, whether it can wait for a batch. Price it on the LLM cost calculator across the models you would accept on quality, which you judge by testing your own prompts.
Then price the same token volume as GPU hours for any model you could run yourself, and add the cost of the people who would run it. The cheaper column under your constraints is the answer, whatever the licence of the weights.
Revisit the answer when volume changes. A workload too small to keep a card busy today may fill several next year, and the column that was cheaper at the start can swap places with the other as traffic grows.
Why the comparison is framed by workload
Comparing models by their headline rate assumes they would use the same tokens for the same job. They often do not. Different tokenizers turn the same prompt into different counts, and different models write answers of different lengths for the same instruction.
So cost per workload is the honest frame. Run a sample of real tasks on each candidate, record the input and output tokens each one used, and multiply by that model's own rates. The result is what that job costs on that model, which is the figure a budget needs.
This frame works the same for any model the data can price and never needs to know the licence, which stays a question for the model's own documentation. Keep the sample tasks, so the same comparison can be rerun when a rate changes.
What we could not answer
- ⚑The data cannot say which models are closed, so no comparison by licence is made here.
- ⚑Self-hosting throughput per card is not something we hold, so the break-even is a method, not a figure.
What this rests on
infratoapp, “Open or closed models: what price data can and cannot say”, read 27 October 2026. models with a token price on record: 467. https://www.infratoapp.com/blog/open-or-closed-models-what-price-data-says
Questions people ask about this
Are open-source LLMs cheaper than closed ones?
Not as a rule. A hosted open-weight model sold per token is priced like any other, and the rate bundles hardware, serving and margin. Open weights can lower cost when you self-host on cards that stay busy, but idle hardware is billed by the hour and someone has to run it.
Can this site tell me which models are open weight?
No. The price data does not record licence reliably, so it cannot say which models are closed. Check each model's own licence page. What the site can show is each model's published token rates by seller, and what a stated workload costs at each.
When does self-hosting an LLM make sense on cost?
When your traffic keeps the GPUs busy most of the time, so the hourly rate is spread over many tokens, and you can staff the operation. Compare GPU hours times the hourly rate, plus engineering time, with your token volume times the API rates for the same workload.
These render as FAQPage structured data · answers are filled from the same live figures
Read next
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.
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.