AI call center agents typically cost in the range of about 5 to 25 cents per minute on managed voice AI platforms, with the wide range driven by the choice of speech-to-text and text-to-speech models, the underlying LLM, and call volume discounts, so any number quoted without your actual usage pattern should be treated as a rough guide only. Pricing usually stacks several components: the LLM's token cost for the conversation, per-minute speech-to-text and text-to-speech charges, telephony minutes through the SIP trunk or provider, and the platform's own orchestration fee if you use one. As of 2026, verify current pricing directly with vendors, since speech model costs have been falling and rates vary meaningfully between premium low-latency voices and standard ones. At meaningful call volume, self-hosting the speech and language models on owned or dedicated GPU infrastructure can reduce the effective per-minute cost well below managed SaaS pricing, though it requires upfront infrastructure investment. The more useful comparison for budgeting is cost per resolved call rather than cost per minute, since a shorter, well-designed conversation can cost less overall even at a higher per-minute rate. Nanobase AI, an NVIDIA Inception Program member, models this full cost stack against a client's call volume before recommending managed or self-hosted GPU infrastructure.
A single per-minute number hides where your money actually goes
Vendors quote a per-minute rate as if it were one price, but that number is really five stacked costs that move independently, and a business that only negotiates the headline rate often misses the components that actually determine their total bill. Understanding the cost stack, not just the final per-minute figure, is what lets you negotiate effectively and forecast spend as call volume grows.
The five components of a voice AI minute
| Component | What drives it up | What drives it down |
|---|---|---|
| LLM tokens | Longer conversations, more context per turn, larger model | Shorter, well-scoped conversations, smaller or self-hosted model |
| Speech-to-text | Premium low-latency streaming tiers | Standard-latency or self-hosted transcription |
| Text-to-speech | Premium natural voices, longer responses | Standard voices, concise agent responses |
| Telephony minutes | Carrier and geography-specific rates | Volume commitments, SIP trunk provider choice |
| Platform orchestration fee | Managed platforms bundling all of the above | Self-hosted pipeline removes this layer entirely |
As of 2026, verify current pricing directly with vendors for each of these components separately, since bundled managed-platform quotes often obscure which piece is actually driving cost at your specific call pattern. Pricing out each component separately, rather than accepting one bundled per-minute number, is what lets you negotiate or self-host the specific piece that is actually expensive for your call pattern.
Why cost per resolved call beats cost per minute for budgeting
A shorter, well-designed conversation that resolves a request in ninety seconds can cost less overall than a longer conversation at a lower per-minute rate, because total cost is minutes multiplied by rate, and conversation design affects the minutes side of that equation as much as vendor selection affects the rate side. Tracking cost per resolved call, calculated as total voice AI spend divided by calls that reached resolution without human handoff, gives a more actionable number than the per-minute rate alone, since it captures both pricing and conversation efficiency in one metric.
When self-hosting the stack changes the math
- Estimate your monthly call minutes and project them forward six to twelve months.
- Get current per-minute quotes from at least two managed platforms at your projected volume.
- Separately price self-hosted speech-to-text and text-to-speech on dedicated GPU infrastructure, plus a self-hosted or API-based LLM.
- Add the operational cost of running that infrastructure, monitoring and on-call capability included, not just hardware.
- Compare the crossover point: at what monthly minute volume does self-hosting undercut the managed platform quote, and is your projected volume past that point.
Self-hosting rarely makes sense below a meaningful volume threshold because the fixed infrastructure and operational cost has to be spread across enough minutes to beat a managed per-minute rate, but past that threshold the gap can be substantial.
Frequently asked questions
Is there a typical range we should expect to pay per minute?
Managed voice AI platforms have historically priced in roughly the 5 to 25 cent per minute range depending on model and voice tier, but treat any number quoted without your actual usage pattern as a rough guide only, and verify current 2026 pricing directly with vendors.
Does a longer call always cost more?
Generally yes in raw minutes, but a longer call that resolves an issue on the first attempt can still be cheaper overall than a short call that fails and requires a callback or human escalation, so total cost per resolved issue matters more than call length alone.
Can we mix managed and self-hosted components?
Yes, some teams self-host the LLM for cost control while using a managed platform for speech-to-text and text-to-speech, or the reverse, matching each component to whichever approach is most cost-effective at their volume for that specific piece.
How much does call volume affect our negotiating position with vendors?
Meaningfully; most managed platforms offer volume-based discounts, so getting a firm monthly minute estimate before negotiating gives you leverage that a vague usage estimate does not.
How Nanobase AI helps
Nanobase AI, an NVIDIA Inception Program member based in Silicon Valley, models this full cost stack against a client's actual call volume and conversation patterns before recommending managed platform pricing or self-hosted GPU infrastructure. This modeling connects directly to sizing a custom chatbot or voice build and to the broader own GPUs versus cloud API cost comparison.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.