What does a completed task actually cost?
Per-token pricing tells you what an attempt costs. It doesn't tell you what a finished piece of work costs.
Explain
A cheaper model with a lower success rate escalates its failures to a dearer one. This tool prices four sourcing options per successful task, using published rate cards and your own volumes. It runs in your browser; nothing you enter is sent to us.
Your workload
Explain
One class is enough to start. Add another if your traffic isn't one shape. Every pre-filled value is an editable illustration, not a benchmark — change anything that doesn't look like your workload.
Committed-managed: the price of a guarantee quote-gated
Reserved capacity bills for every minute of the term whether you use it or not. This is the guarantee ladder, not a GPU-reservation ladder.
Explain
Traffic above the reservation spills to serverless at list, best-effort, outside the SLA. So the question isn't whether it's cheaper — at this provider, on current data, it is never cheaper than serverless at any utilization, and equal only at full use — it's what covering the traffic you can't afford to have throttled costs you.
Serverless managed-open: what the rate cards say
Self-host: undetermined until measured
Explain
Self-host cost depends on how efficiently your serving stack turns paid GPU time into useful tokens. That is set by your configuration and cannot be read from anything you can type. The tool shows self-host as a band, and shows a number only when every input is yours.
Where the numbers come from
Take this with you
Explain
Nothing here has been stored. If you'd like a copy, the first button opens a pre-filled email on your own device and the second copies the same text; if you'd like to talk it through, the third takes you to the contact page.