Pricing & billing
TurboPrivate uses prepaid USD credit and settles completed model work from measured input, cached-input and output tokens. There is no customer charge for an idle room.
Token prices
| model | input /M tokens | cached input /M | output /M |
|---|---|---|---|
Gemma 4 26B A4B gemma-4-26b-a4b | $0.50 | $0.50 | $2.00 |
Qwen3.8 27B qwen3.8-27b | $1.50 | $1.50 | $6.00 |
Prices are per million tokens. The CLI and SDKs receive these same values from the product catalog.
Decode speed and elapsed time do not change the unit price. Any “per turn” number in a model picker is an example based on a stated token mix, not a flat fee or a promise about the final charge.
How a turn is charged
- Before upstream work, the room reserves a conservative maximum against the account's globally available balance.
- The model returns usage counters. TurboPrivate recomputes cost from the selected product's catalog prices.
- Settlement debits measured usage and releases the unused reservation.
A reservation is a temporary hold, not a charge. It prevents concurrent rooms from spending the same credit. If the available balance cannot cover the safe maximum for a turn, the request stops with insufficient balance before model work begins.
Cached input
When the engine reports repeated prompt-prefix tokens as cached, those tokens use the cached-input price in the table, which can equal the ordinary input price. Cached tokens are part of total input, not an additional charge. Cache availability is not guaranteed, and applications should budget at the ordinary input rate. See Context & caching for request examples and usage fields.
Streaming and cancellation
A client that cancels a stream may do so before the engine emits final usage. In that case the room settles a conservative local estimate and records it as estimated. It does not charge the full preauthorization merely because final counters were unavailable.
Balance and top-ups
Open Balance in the console to see credit and usage. Choose Card or Stablecoins, then choose the top-up amount; Stripe Checkout opens with the selected payment rail. Payments are processed by Stripe; TurboPrivate receives the confirmed amount and does not store the full card number.
When Stripe marks its Stablecoins method as available, a customer can pay from a compatible wallet with a supported USD stablecoin—USDC, USDP or USDG, depending on network and eligibility. Stripe settles the confirmed payment in USD and TurboPrivate adds the same USD amount to the prepaid balance. This checkout does not accept Bitcoin (BTC), Ether (ETH) or other volatile crypto assets.
Stablecoin payments are not anonymous cash: Stripe processes the payment and the transfer is visible on a public blockchain. TurboPrivate stores the Stripe session identifier needed for idempotent crediting, not a customer wallet profile.
Unused prepaid credit is refundable on request within 30 days of purchase. For a failed session, billing error or unexpected charge, use the support contact on the refund policy.
Model routes
All inference is requested and billed as an approved TurboPrivate catalog model running on hardware TurboPrivate operates. Customer traffic is not routed to third-party model providers.
Budgeting correctly
- Use the selected model's input, cached-input and output prices rather than an editor's built-in vendor estimate.
- Budget uncached input unless your workload has measured cache behavior.
- Higher reasoning effort can add output/reasoning tokens and first-token latency.
- Opening more sessions increases possible concurrent spend; the global reservation gate prevents balance overcommit.