Pricing
Inference draws on prepaid credits, billed one of two ways. A catalog model bills the tokens each request uses. A model you deploy bills the GPU time it holds. Deploying is free.
Catalog models
The open models Boardwalk serves are billed per token, at rates quoted per million tokens. There is no minimum usage or reserved-capacity fee. The model catalog and GET /v1/catalog/models show the active rates.
Rates are fixed for a configuration revision. A new price requires a new revision; past receipts keep their original rates. Traffic spikes do not change the price of a revision.
Token rates
| Category | Billed tokens |
|---|---|
| Input | Uncached prompt tokens at the input rate. |
| Cache read | Prompt tokens served from the engine’s prefix cache. |
| Cache write | Tokens the engine reports writing to its cache. Free under automatic prefix caching. |
| Output | All generated tokens, including reasoning, at the output rate. |
If the engine does not report cache counts, the full prompt is billed as input and the receipt records that cache usage is unknown.
Your deployments
A model you deploy bills GPU time, not tokens. You pay per card, per millisecond, from the moment a card is granted to your model until it is released. There is no minimum.
| Card class | On demand, per card-hour | Reserved capacity, per card-hour |
|---|---|---|
| H100 | $4.50 | $3.25 |
| H200 | $5.75 | $4.25 |
The console lists the rate of every class, B300 included. A replica of several cards bills each card. A slice of a card bills its share of the card’s rate. An on-demand rate is fixed when the card is granted: no surge pricing. Reserved capacity follows the policy below.
| Billed | Never billed |
|---|---|
| The swap onto the card (the cold start), serving, and the hold | Deploys, engine boots, staging, the copy in host memory, waiting for a card, and any visit that fails for a reason of ours |
The hold is the idle time after the last request before the card is released: 1 to 300 seconds per model, 5 by default, always honored. Requests running at once share the card while its engine has room. A request carries no per-token price; its receipt records token counts at zero rates and links the GPU time that paid for it.
swap 0.3 s + serving 42.0 s + hold 5.0 s = 47.3 s
47.3 s × $5.75 / 3600 = $0.0755Reserved capacity
Reserve individual GPUs, groups of 2 or 4, or full 8-GPU nodes. Each group stays together on one host. Capacity stays enabled until you reduce its quantity to zero. Your models share these GPUs by swapping in; traffic beyond your reservation uses on-demand GPUs.
- Bills per millisecond from allocation, even while idle, with a one-hour minimum for each GPU added. GPUs awaiting allocation cost nothing.
- Edit the quantity of each GPU class and group size. Reductions show as releasing until the GPUs are released; active requests continue at their on-demand rate once their GPU is no longer reserved.
- An admin enables groups through New capacity under Capacityin the console. Each GPU class and group size has one row with its quantity, allocation status and current hourly cost. Only GPU classes offering a reservation rate are available.
- New capacity uses the published rate immediately. Existing capacity receives at least 24 hours’ notice of increases; decreases apply automatically. The capacity page shows upcoming prices and when they take effect. Earlier charges keep their recorded rates.
- Reserved and on-demand GPUs share a capacity limit based on your highest credit balance: $50 allows 8 GPUs, $250 allows 16, $1,000 allows 32, and $5,000 allows 64. Spending credits never lowers the limit.
Stored weights
Each copy of weights your deployed models keep in Boardwalk’s store bills per GB-month, accrued hourly, whether or not it is called. A copy is your organization’s own: another organization’s copy of the same files never covers yours.
Add credits
Open Credits in the console to purchase credits and view your balance. Purchases start at $5, and credits never expire. New organizations start at zero, and nothing is served before the first purchase. The organization making an API call pays for it, including calls to public models.
Auto top-up is off by default. Enable it to charge the card from your last purchase when usage leaves your balance below a chosen threshold: a request, or a deployment’s GPU time as it settles each minute. If the balance is insufficient, new inference requests return 402 and no new card is granted; reservations end at the close of their hour unless you add credits. Read-only organization data remains accessible.
Track spending
Charges use exact integer micro-dollars; a token charge is rounded once on the request total. The console shows usage by model and individual request receipts. Each receipt records the tokens, exact rates, total cost, and serving revision.
A catalog request that fails or disconnects costs nothing until output reaches you; after that it bills what was generated. On your own deployment, the card’s time is billed whether a request succeeds or not. Requests refused before reaching a model are free. Tokenization is free.
See metering and receipts to verify a charge, or use the Requests API to export request history.