For contact-centre operators
At 50,000 minutes a month and up, the per-minute meter stops being a convenience and starts being the budget. PeachDesk gives operators a way off the meter: run inference on your own GPUs, keep cost and margin visible on every call. This page is part of our voice AI solutions overview.
Per-minute pricing reads well in a pilot. The same rate at contact-centre volume punishes success: every deflected call adds minutes, every added minute adds cost, and the provider rate is a floor you cannot negotiate away because the platform owns the model plane. Your unit cost is a markup on someone else's infrastructure decision.
Owning the inference plane changes the shape of the cost curve. In Local mode, speech-to-text, the language model, and text-to-speech run as a full open-model stack on your own GPUs: no per-minute provider charge on self-hosted stages, and you pay for GPU capacity. Cost becomes a capacity decision. As utilization rises, inference cost per minute falls, instead of holding flat at a vendor rate. Frontier mode stays metered for the stages where you want leading cloud providers, and telephony carriers always meter; you mix sourcing per pipeline stage, per agent.
Not every operation wants to run GPUs. The managed Bibha Plan meters at $0.06 per minute, as low as $0.03 per minute at volume, and teams on their own provider keys can start with the $59 per month Pay As You Go plan. The full breakdown is on the pricing page and in our explainer on voice AI cost per minute.
Cost ownership only counts if the platform runs the operation. The pieces contact-centre teams ask about first:
The right structure depends on your volume, your team, and whether you already operate infrastructure. The mechanisms, not quotes:
| Model | Cost floor | Who it fits |
|---|---|---|
| Pay As You Go | $59 per month for teams on their own provider keys; providers bill you directly for usage. | Teams validating voice AI on existing provider accounts before committing volume. |
| Bibha Plan (managed) | $0.06 per minute, as low as $0.03 per minute at volume. | Operators who want a managed stack with a rate that improves as committed volume grows. |
| Self-hosted with Local mode | Your own infrastructure plus GPU capacity. No per-minute provider charge on self-hosted stages; you pay for GPU capacity. | High-volume operations where owning the inference floor beats renting it, and where deployment ownership is already part of the model. |
Billing is off by default in the self-hosted stack; you still pay your own providers and infrastructure. What disappears is the vendor in the middle of every minute.
PeachDesk has not published load-tested benchmarks, and we will not quote concurrency or throughput numbers we have not measured. What we offer instead is a guided proof-of-volume pilot on your own infrastructure: we size GPU capacity against your real call profile, run your traffic, and measure cost per minute under your load. You make the platform decision on your numbers.
Frequently asked questions
Because the per-minute meter scales linearly with volume while your budget does not. At 50,000 minutes a month, every point on the provider rate is a real line item, and you cannot lower the floor because the platform owns the model plane. PeachDesk gives you a way off that floor: Local mode runs inference on your own GPUs with no per-minute provider charge, and you pay for GPU capacity.
In Local mode, speech-to-text, the language model, and text-to-speech run as a full open-model stack on your own GPUs. There is no per-minute provider charge on self-hosted stages; you pay for GPU capacity, so inference cost becomes a capacity decision you control rather than a meter that grows with every call. Frontier mode stays available for the stages where you want leading cloud providers, mixed per pipeline stage.
The managed Bibha Plan meters at $0.06 per minute, as low as $0.03 per minute at volume, so the rate improves as your committed volume grows. Teams that already hold provider keys can start on the $59 per month Pay As You Go plan on their own keys and move to self-hosting when the volume justifies it.
No, and we will not invent them. We have not published load-tested concurrency or throughput benchmarks. What we offer instead is a guided proof-of-volume pilot on your own infrastructure: we size GPU capacity against your real call profile and measure cost per minute under your load, so you make the platform decision on your numbers, not our slide deck.
Bring your volumes, your queues, and your carrier contracts. We will walk through the cost structure on the operation you actually run.