Ollama switched its hosted plans from GPU-time billing to per-token pricing

Ollama posted a new pricing structure for its hosted service on 31 August 2026. Pro, Max and Team plans now bill in tokens against a published rate rather than in GPU time, and each plan comes with a fixed monthly credit pool. Existing subscribers keep the plan they signed up on and can switch through account settings when they want to.

The reason Ollama gives in the post is direct: "GPU-time based billing was difficult to predict, especially as open models have grown much larger (Kimi K3 has 2.8 trillion parameters)." A larger model consumes more GPU time per token, so a session's cost varied with the model as well as with the length of the answer, and users could not tell in advance which one was driving the bill.

The new numbers

Pro is $20 per month and comes with $60 of monthly credits at the published per-token rates.

Max is $100 per month and comes with $300 of monthly credits.

Team is $500 per month, carries $1,000 of shared monthly credits, and covers unlimited users on one account.

The Free plan keeps a small monthly allocation limited to starter models and adds pay-as-you-go on top for anyone who wants to test a bigger model without a subscription.

Two conditions are worth reading twice. Unused credits do not carry over: Ollama's post says "your included amount refreshes at each monthly reset", so credit not used in a quiet month cannot be used in a busy one. And the per-token rates are the same across every plan, so the monthly fee decides the size of the credit pool, not the unit price of a token.

Tokens replace GPU time as the unit

If you already use Ollama Cloud, the immediate work is a spreadsheet check. Pick your two or three heaviest workloads, apply the published token rate to the input and output tokens they actually produce in a normal week, and see whether $60, $300 or $1,000 of monthly credits covers them. The old GPU-time billing hid the answer inside a variable session cost, so anyone who dismissed Ollama as unpredictable can now put a real number on it.

Two structural implications follow. Cost of any specific model on Ollama Cloud is now directly comparable to the same operation on OpenAI, Anthropic or a Together AI endpoint, because both sides quote a per-token rate. And a Team account for $500 a month with unlimited seats is a fixed cost, which is easier to defend than seat-based licensing for a small engineering group that runs experiments at irregular times.

The problem is the migration path for existing subscribers. Ollama says the old plans stay live for anyone already on them, which is fair, but it also means anyone budgeting for the next quarter has to choose between the old and new plans. Nothing in the post suggests the old GPU-time plans are going away, but every future feature will be priced against the new structure.

Source

Primary source: Ollama's transparent pricing, Ollama blog, 31 August 2026.