Atlas
One ledger for every model
Atlas meters self-hosted Whisper by the word and every other model by the token — under one registry that never runs out of room for the next one.
Two pricing strategies, one system
Every model in Atlas maps to exactly one strategy. Adding a model never means adding a special case.
Self-hosted Whisper — per word
Transcription cost tracks what the model actually produced: words on the page, not tokens under the hood. Cost = words transcribed × rate for that Whisper variant.
Every other model — per token
Language, vision, and everything else bills on tokens consumed, the native unit those models already report. Cost = tokens used × rate for that model.
No ceiling on models or strategies
The registry holds the models. The strategy decides how they're priced. Neither one is hardcoded against the other.
Add a model
One registry entry: id, strategy, rate. No code change, no redeploy.
Add a strategy
Per-second audio, per-image, per-call — one new strategy module, existing models untouched.
Usage stays flat
The usage ledger only ever stores units consumed and cost — schema never migrates for a new model.
Every model, one table
| Model | Category | Strategy | Rate |
|---|---|---|---|
| Whisper Tiny | Self-hosted transcription | Per word | 1.5 tokens / word |
| Whisper Base | Self-hosted transcription | Per word | 2.5 tokens / word |
| Whisper Small | Self-hosted transcription | Per word | 4 tokens / word |
| Whisper Medium | Self-hosted transcription | Per word | 5.5 tokens / word |
| Whisper Large v3 Turbo Q5_0 | Self-hosted transcription | Per word | 6 tokens / word |
| Whisper Large v3 Turbo | Self-hosted transcription | Per word | 7 tokens / word |
| Whisper Large v3 | Self-hosted transcription | Per word | 9 tokens / word |
Meter anything, forever
Atlas doesn't care how many models you add next year. It only cares which strategy they use.
Back to top