Ollama put a weekday afternoon surcharge on open models
New per-token pricing gives Pro subscribers sixty dollars of usage for twenty, and doubles DeepSeek's rate for six hours of every weekday.
Ollama has put a rush-hour surcharge on open-weights models.
On 1 September it moved its Pro, Max and Team plans to per-token pricing, with a published rate card for every model it hosts. Pro is $20 a month and includes $60 of usage. Max is $100 and includes $300. Team is $500 and includes $1,000, shared across unlimited users.
The company whose name is a byword for running a model on your own machine now sells inference by the token, with concurrency tiers and a peak rate.
Six hours, two models
Peak pricing applies between 12:00 and 18:00 UTC, Monday to Friday, and it applies to exactly two models. The DeepSeek Flash model goes from $0.22 per million input tokens to $0.44. The Pro model goes from $0.66 to $1.32.
Both double. Nothing else on the list carries a peak rate at all.
Six hours on weekdays is roughly the European afternoon stacked on the American morning, which is probably where the demand curve for coding agents lives. And the fact that only the DeepSeek family carries the surcharge tells you where capacity is tight, which is the sort of thing an inference host normally keeps to itself.
That is a spread of more than forty to one, across models a user picks off the same menu.
What the included credit actually buys
The rate card arrived days after GLM-5.3 and its Flash sibling finished rolling out on the cloud, where they are available cloud-only, with no downloadable build.
Flash sits at $0.15 per million input tokens and $0.50 output, the larger model at $1.40 and $4.40. So a Pro subscriber's credit buys roughly four hundred million input tokens of the small one, or about forty-three million of the big one. Input only, mind, and cached input is cheaper again.
Concurrency is capped too: one request at a time on the free tier, three on Pro, ten on Max and Team.
The plan aimed at companies is the worst value on the page
Read the tiers as a subsidy and the shape is clear. Pro turns twenty dollars into sixty of usage, Max turns a hundred into three hundred. Three to one, both times. Team turns five hundred into a thousand, which is two to one.
That is normal in software, and slightly funny on a price list that opens with the words "no service fees or hidden limits." Unused credit does not roll over either, so three to one is a ceiling rather than a balance.
Open weights now mostly means choosing your landlord
This is the moment the open-weights story stopped being about your own GPU. Ollama hosts primarily in the United States, routes to Europe and Singapore for extra capacity, works with NVIDIA Cloud Providers, and requires no logging, no training and zero retention from them.
Those are the commitments of a utility, not of a download.
Does that make it worse? Not obviously. A rate card you can read beats a subscription whose limits you discover by hitting them, and Ollama is unusually blunt about what things cost (the peak table is right there on the page, and it did not have to be).
But if you moved to open models to stop depending on somebody else's capacity planning, check the clock before your next long agent run.
We would expect the peak window to spread beyond DeepSeek to at least one more model family before the end of the year. If it does not, Ollama will have quietly solved a capacity problem the rest of the market is still pretending it does not have.
