If you run open-weight models on Ollama’s cloud, the way you pay changed on the last day of August. The company replaced its GPU-hour billing with a flat monthly subscription that comes with a token-credit pool, topped up at the same per-token rate once you spend it down. For a user who already had a plan, nothing changes until you upgrade, but every new signup, and every plan that flips after the announcement, now prices the same way the rest of the industry does.
That last part is the story. GPU-hour billing is the kind of meter that punishes long contexts and rewards short, cheap ones, and as Ollama’s post puts it, the company “received feedback that GPU-time based billing was difficult to predict, especially as open models have grown much larger (Kimi K3 has 2.8 trillion parameters).” Predictability, not lower prices, is the headline.
What the Tiers Actually Cost
Ollama’s pricing page lays out five plans, of which four are individually priced and Enterprise is “custom.”
The free plan stays free, with “a starter amount each month for a set of starter models,” and an optional pay-as-you-go top-up “with no service fees.” Pro is $20 per month with $60 of monthly usage included. Max is $100 per month with $300 of monthly usage. Team is $500 per month with $1,000 of shared usage “for unlimited users” drawn from one pool. Each plan refreshes on the same day of the month it started; unused credits do not roll over.
Concurrency used to be a tight constraint: Ollama’s August blog post explicitly drops “no 5-hour or weekly limits” as a feature, and the pricing page now publishes concurrent-request ceilings of 1 for Free, 3 for Pro, and 10 for Max and Team. Overage continues at the same per-token rate once the included pool is spent, with an email alert at 90 percent on paid plans.
The interesting numbers are per-model. Per the Ollama pricing page, gpt-oss:20b runs $0.07 per million input tokens and $0.30 per million output. gpt-oss:120b runs $0.15 and $0.60. gemma4 is $0.14 input, $0.40 output. qwen3.5:397b is $0.60 input, $3.60 output. deepseek-v4-flash is $0.44 input, $1.32 output. The biggest open-weight model listed, kimi-k3 at 2.8 trillion parameters, is $3.00 input, $15.00 output per million tokens.
Those are real numbers a reader can run with. A prompt-plus-response workload averaging 1,000 input tokens and 500 output tokens on qwen3.5:397b costs roughly $0.0024 per call at list (1,000 × $0.60 per million input plus 500 × $3.60 per million output). The $60 Pro pool buys about 25,000 such calls per month at list; in practice, real workloads mix cheap and expensive models and the pool is a guardrail more than a budget.
What Stays the Same and What Disappeared
Three things stay. The desktop experience, where you ollama run a model locally against your own GPU, is unaffected and still free. The Ollama library of open-weight models stays the same, with Ollama’s August post explicitly emphasising that it is “compatible with Claude Code and Codex coding agents,” which means you can point those clients at Ollama’s hosted endpoints the same way you would point them at OpenAI or Anthropic.
Three things changed. The old GPU-hour meter is gone, replaced by per-token accounting. The 5-hour and weekly caps that defined the original Pro and Max plans are gone, replaced by a monthly credit pool. And the data-handling story is now stated positively rather than as an absence: per the Ollama blog, “We don’t log your prompts, and we never train on your data,” with compute described as “dedicated” rather than shared multi-tenant.
Hosting sits in the US and Europe, with Singapore capacity for a limited set of Qwen models per Ollama’s August blog post. The terms on those pages are also unusually explicit about no training on user prompts, which is the practical question for a privacy audience running on a hosted model.
Why This Matters Beyond the Numbers
Ollama is now the largest open-weight runner by developer count. The Next Web reported that the company closed a $65M Series B on 9 July 2026 led by Theory Ventures, with participation from Benchmark, 8VC and Y Combinator, taking total funding to $88M; the same piece cited “8.9 million monthly active developers” and presence in “85 percent of Fortune 500 companies.” That is the scale that turns a billing change into an industry signal.
Benchmark partner Peter Fenton, who sits on Ollama’s board from the earlier round, said in the funding announcement: “Open-weight models will generate the supermajority of tokens within the next 18 to 24 months.” Read together with Ollama’s choice to drop GPU-hour billing and align to per-token rates, the prediction and the product change point the same way: the open-weight camp wants to be priced like the closed-weight camp so that procurement teams can compare line items.
The privacy story is the part to watch most carefully. Ollama’s no-logging, no-training claim is a self-published promise, not a third-party audit, and the pricing page does the right thing by stating it in plain language: “We don’t log your prompts, and we never train on your data” and “Every request runs on dedicated compute.” That is stronger language than the major closed-weight APIs publish, but it is also unaudited, and a reader who needs the strongest possible guarantee still has to run their own models on hardware they control.
What This Means
For a developer who already runs Ollama’s cloud, the immediate change is small: existing plans keep their terms until you opt in, and the new tiers are strictly more generous in concurrency and predictability. The interesting choice is whether the new Max at $100 with $300 of credits makes sense, which depends entirely on the model mix. A user who lives on gpt-oss:20b at $0.07 per million input will barely register spend on the $60 Pro pool. A user who runs kimi-k3 at $3.00 per million input will exhaust that same pool in roughly 20 million input tokens, which is not a lot of long-context work.
For the local-AI ecosystem, the change closes the last operational gap between Ollama’s cloud and the major closed-weight APIs: predictable per-token pricing, no daily cap, and Claude Code / Codex compatibility out of the box. That matters for any team evaluating open-weight versus closed-weight on cost grounds. It does not change the privacy calculation, which still comes down to whether you trust Ollama’s no-logging claim versus running on your own hardware.
The Bottom Line
Ollama’s cloud now prices like everyone else’s: a monthly subscription plus a per-token credit pool, with a free tier that stays free. The change is welcome for predictability, neutral for cost if your workload mixes models, and worth a hard look if your workload is heavy on the 2.8-trillion-parameter Kimi K3. Privacy-wise, take the no-logging promise at face value until an audit says otherwise, and keep a local install in the toolbox for anything you would not send to a hosted endpoint under any terms; our Ollama and Open WebUI self-hosting guide walks through the 15-minute setup.