What it means, and what it does not
An open-weight model means the vendor publishes the trained weights: the several billion numbers that are the model. It can be downloaded, run on your own hardware, modified. What exactly separates this from closed models is the subject of Open Weights vs Closed Models, a separate article.
This article is about something else: about why it matters even if you never download anything. It has five effects, and you feel all five on your invoice and in your contracts.
The first effect: a price ceiling
This is the most concrete one. In 2026 the cheapest serious open-weight models run at around $0.14 and $0.28 per million tokens for input and output. The same figures at top-tier closed models are $5–10 and $25–50.
On the output side that is orders of magnitude cheaper — one public analysis put a particular comparison at roughly 150 times. The same model is not suitable for every task, but it is for a substantial share of them.
The effect of that pressure is visible: in the last two years every large closed provider has introduced a cheap tier, a caching discount and batch runs at a reduced rate. Those did not all appear at the same time by accident.
The second: the size of the gap
The usual counter-argument is that open models are weaker. That is true — but the question is by how much.
The current estimate is roughly three to six months behind the best closed models, and that gap has not been widening. Translate that into practice: if an open-weight model today does what the closed frontier did six months ago, then for most everyday tasks it is more than enough — because the closed frontier solved those six months ago too.
The third: the EU exit
This is why the topic comes up at European companies at all. If the weights are downloadable, the model can run inside your own boundary: on your own servers, in your own data centre, or at an EU provider you contracted with yourself.
In that case the data location is not a promise but a fact — you do not have to believe where it runs, because you are running it. That does not solve legal compliance by itself, it only takes the hardest part out of it; the rest is what European AI Options deals with.
The fourth: nothing gets retired
Closed models get retired. Vendors generally give notice, but the date arrives, and the prompts tuned to that model will behave differently.
A downloaded model cannot be retired. If you saved it, it will be exactly the same model in five years — which for audit, reproducibility and long-lived systems is not a small thing. Your Model Was Deprecated goes into that in more depth.
The fifth: bargaining position
This is the least visible effect and the most valuable. If a working alternative exists, then there is something to negotiate against.
A provider who knows you have nowhere to go prices the renewal differently from one who knows you do. You do not have to switch — it is enough that you could. This is the same logic as in any supplier relationship, except that here the alternative is a free download.
What the word "open" does not mean
It is worth being precise here, because the wording misleads:
- Not open source. The weights are public; the training data and often the training code are not. "Open-weight" is a more accurate phrase than "open source".
- Not always free to use. Licences differ: some releases are MIT-licensed, others restrict by user count or by purpose. Before commercial use, this has to be read.
- Not free of cost. The weights are free, running them is not. With self-hosting, the price of hardware, electricity and an engineer replaces the price of tokens — at low volume that is typically more expensive, not less.
What to do with this if you never run one?
Three practical consequences, all three usable immediately:
- Include it in the comparison. When you choose a model, look at the open-weight line too — through a router it is reachable without hosting anything.
- Mention it at renewal. Not as a threat but as a fact: an alternative exists, and you know it.
- Try it on the bulk work. Classification, extraction, tagging — on these tasks the difference is often unmeasurable, while the price difference is not.
What to take away
Open-weight models work for you even if you never run one: they hold prices down, they offer an EU exit, they cannot be retired, and they create a bargaining position. Their gap is around six months and is not growing — which is plenty for most everyday tasks. In exchange, "open" does not mean free to use, and it does not mean free to run.