Costing path · Serving choice
Serving choice
Pay-per-token or dedicated capacity for the same volume. Dedicated cost rises in whole copies, so the answer can flip; the break-even is the volume above which self-hosting stays cheaper, worked out exactly rather than by sampling.
Jump to results
Result
Recalculates as you type
At 10 bn tokens a month, pay-per-token is cheaper. Self-hosting is cheaper at every volume above 24.92 bn tokens a month. Between about 18.8 bn and that point it depends on the exact volume: each extra GPU copy adds a step in cost, so self-hosting can be cheaper at one volume and dearer at the next.
Pay-per-token, monthly
$6,000.00
$0.60 per million tokens
Self-hosted, monthly
$7,650.00
$0.77 per million tokens, including amortised one-offs
Self-hosted detail
- Copies needed
- 1
- Utilisation of those copies
- 50%
- Monthly run cost
- $7,650.00
- First-month outlay
- $7,650.00
- Cheaper at every volume above
- 24.92 bn tokens
- Can be cheaper above
- About 18.8 bn tokens, depending on the exact volume
Monthly cost lines
| Cost line | Monthly |
|---|---|
| Pay-per-token: token chargesModel | $6,000.00 |
| Self-hosted: GPUs (1 copy)Service | $3,650.00 |
| Self-hosted: people running itPeople | $4,000.00 |
| Self-hosted: migration and evaluation, amortisedBuild, amortised | $0.00 |
Use in a workflow
Not ready to present as a result
- Capacity per copy is illustrative, not measured or quoted. Do not present this comparison as a result until capacity is measured or quoted.
Limits of this estimate
- Planning estimate: steady volume across the month; one blended token price; capacity is per copy at full load.
- Illustrative pay-per-token price.
- Illustrative GPU rate.
Save and share
The JSON re-imports exactly and records the methodology version. It carries the baseline cost when one is entered. Files are made in your browser; nothing is uploaded.