Buy any amount
There are no packs to choose between. Drag to whatever you want to spend — anything from €10 to €100 — and top up again whenever you like.
- ·Credit never expiresNo monthly reset. A quiet month is a month you did not spend.
- ·No seats, everAdd the whole team. You are charged for runs, nothing else.
- ·One published rateEvery run costs 1.6–1.8x what it cost us to produce, €0.01 floor, for everyone.
- ·Stop wheneverNothing to cancel. Stop topping up and you stop paying.
How model usage is billed
One thing is being sold: model usage. Everything around it — agents, versions, diffs, audit, per-agent access control, as many teammates as you like — comes with it at no extra charge.
You buy credit from us
Coldcache holds the accounts with Anthropic, OpenAI and Google. You top up a balance and each run draws down against it. One bill, no keys to rotate, no provider account to open, and nothing to cancel — stop topping up and you stop paying.
1.6–1.8× the run cost
Priced off the real cost of the run — tokens, thinking and every web search it made — with a €0.01 floor, and the multiple drops as you top up more. The run record shows what it cost us and what you were charged, side by side. A rate change never rewrites an old invoice, because the charge is stored at execution time.
Your own provider keys
If you have committed spend with a provider, or a policy that says model calls leave from your account, Enterprise can run on your keys instead. It is enabled per workspace by us, and the platform is then licensed directly rather than paid for through credit.
Questions people actually ask
Is there a per-user fee?
No. There are no seats, no monthly minimum and no subscription. Add your whole team, give them access to whichever agents they need, and you are charged for exactly one thing: the runs they produce. A teammate who reads other people's threads all day and never runs anything costs you nothing.
What counts as a run?
One answer generation, including every tool call and web search it makes along the way. A streaming answer you cancel before it finishes is not charged. Draft test-drives in a scratch chat are not charged either, and a run we refuse — out of credit, suspended — is never charged.
How is a run priced?
At 1.6 to 1.8x what the run actually cost us to produce, depending on your top-up size, with a €0.01 floor. The cost is real token usage, not an estimate, so a one-line answer costs a fraction of a cent and a deep research run with thirty searches costs more. Every run record shows what it cost and what you were charged, side by side.
Does credit expire?
No. What you buy stays on the balance until runs spend it. There is no monthly reset, so a quiet month is simply a month you did not spend anything.
What happens when credit runs out?
Runs stop and nothing else changes. Agents, versions, threads and every run record stay exactly as they were — nothing is deleted, downgraded or locked for a zero balance. You get a warning at €5 remaining and runs resume the moment you top up.
Can we use our own OpenAI or Anthropic keys?
On Enterprise, yes. Everywhere else, no — runs execute on Coldcache's model access and you buy credit from us. It is not a technical limit, it is how the product is sold: one bill, no key rotation, and a per-run cost record we can stand behind. If you have committed spend with a provider and want to draw against it, that is an Enterprise conversation.
Do I pay for versions or team members?
No. Agents, versions, version history, threads, diffs, the audit log and per-agent access control are unlimited at every level including free. Metering the thing that makes the product work would undercut the entire pitch.
Can I cap what a workspace spends?
Yes. Set a monthly run cap per workspace and runs stop at the number instead of drawing the balance down further. Nothing is silently throttled — a slower answer you did not ask for is worse than a clear stop.
What happens to my data when retention expires?
Nothing is hard-deleted. Content is redacted and the hashes, config and telemetry are kept, so a diff can still tell you whether something changed even after the text is gone.
Can I self-host?
On Enterprise, yes — VPC or fully on-premise, with your own database. The versioning semantics are identical; the only difference is where the rows live.
Try it on one real workflow.
Publish v1, run it, change the prompt, diff the two. Four minutes is enough to know whether the pin is worth anything to you.