Cost per Token
Every row joins a supply price to a benchmark throughput figure. The result is USD per million output tokens — the number that decides whether a deployment pays for itself.2,200x spread between the cheapest and most expensive option.
| USD/Mtok | Concurrency | Tier | Provider | Accelerator | Model | tok/s/GPU | $/GPU-hr | Status |
|---|---|---|---|---|---|---|---|---|
| Loading cost data… | ||||||||
How to read this
Aggregate rows are fleet-average costs: the benchmark measured total output tokens/s across all in-flight requests. Single-user rows are what one request sees. These are different populations — never compare an aggregate row to a single-user row.
The tier column matters as much as the price. Spot and peer-to-peer prices are 2-5x cheaper but can be revoked with minutes' notice. On-demand is the only tier you can build a production SLA on.
Every row links to both the supply/ and benchmarks/ records it was derived from. The join is computed by tools/cost_per_token.py — no numbers are hand-entered.