How much does an AI voice agent cost per minute in 2026?

The four-band cost model

The 2026 market sits in four overlapping bands. Self-assembled cascaded stacks run $0.07 to $0.13 per connected minute for production traffic and can drop to $0.007 at the most aggressive committed-use tiers. Speech-to-speech models run $0.18 to $0.21 per minute and buy native audio expressivity with less per-stage control. Developer platforms advertise $0.05 to $0.31 per minute but the loaded cost lands 2x to 3x higher once LLM, STT, TTS, and telephony stack on. Done-for-you solutions charge $500 to $2,000 a month and trade variable rate for predictable billing plus zero engineering overhead.

The right band for a project depends on call volume, expected failure rate, and how much engineering time the team can absorb. The crossover between platform premium and self-assembly tends to land somewhere between 50,000 and 100,000 connected minutes per month, but the exact point depends on team cost, reliability requirements, and traffic shape.

• Cascaded self-assembled stack: $0.07 to $0.13 per minute, can drop to $0.007 at committed-use tiers.
• Speech-to-speech model: $0.18 to $0.21 per minute for native audio expressivity.
• Developer platform headline: $0.05 to $0.31 per minute, loaded cost 2x to 3x higher.
• Done-for-you solution: $500 to $2,000 a month for predictable billing.

What always stacks on top of the headline rate

Telephony is the floor cost on every call and is rarely included in headline per-minute pricing. Real numbers sit around $0.01 to $0.025 per connected minute depending on carrier, call type, and number geography. Inbound calls on a local French number are around $0.01 per minute; outbound calls to a mobile are around $0.04 per minute; a US inbound number carries an additional $1 to $2 monthly fee.

Speech-to-text typically bills per audio minute at $0.015 to $0.030 for streaming STT. LLM tokens are the biggest swing and run $0.020 to $0.080 per minute depending on model choice and conversation length. Text-to-speech is the dominant cascaded cost line and runs $0.020 to $0.060 per minute, with premium or cloned voices at the top of the range. Orchestration and platform fees add $0.005 to $0.020 per minute on platforms that bill them separately.

• Telephony: $0.01 to $0.025 per minute plus number rental.
• Speech-to-text: $0.015 to $0.030 per minute of streamed audio.
• LLM tokens: $0.020 to $0.080 per minute depending on model and history growth.
• Text-to-speech: $0.020 to $0.060 per minute of generated speech.
• Orchestration and platform: $0.005 to $0.020 per minute on platforms that bill separately.

The architecture split: cascaded vs speech-to-speech

Cascaded stacks (separate STT, LLM, and TTS) are cheaper at scale and give granular control. The same 2026 production benchmark shows cascaded deployments at $0.07 to $0.13 per minute with medians around $0.10. The cost advantage comes from being able to swap each component independently: tune STT endpointing for noisy environments, swap LLM per call type, cache TTS for fixed prompts.

Speech-to-speech models handle audio in and audio out through a single model, which skips the hand-offs between separate services and posts the lowest p50 latencies (540 to 580 milliseconds in the same benchmark). The trade is cost ($0.18 to $0.21 per minute) and less control over each stage. Native audio expressivity is the premium you are paying for.

How volume moves the per-minute rate

A typical 2026 voice agent deployment shows three volume bands. Around 1,000 minutes per month pays full retail rates at $0.16 to $0.21 per minute ($160 to $210 monthly). Around 10,000 minutes per month unlocks volume STT and TTS tiers plus model routing and lands at $0.11 to $0.15 per minute ($1,100 to $1,500 monthly). Around 100,000 minutes per month unlocks committed-use discounts plus caching plus smaller models for most turns and lands at $0.07 to $0.11 per minute ($7,000 to $11,000 monthly).

The crossover between buying a per-minute platform and running a self-assembled stack depends on team cost. At modest volume, paying a higher variable rate can still be cheaper overall once engineering time is included. At very large volume, saving a few cents across millions of minutes can justify owning more of the infrastructure.

• About 1,000 minutes per month: $0.16 to $0.21 per minute, full retail rates.
• About 10,000 minutes per month: $0.11 to $0.15 per minute, volume tiers active.
• About 100,000 minutes per month: $0.07 to $0.11 per minute, committed-use tiers active.

Cost per successful outcome beats per-minute cost

The right metric for budgeting is cost per successful outcome, not cost per connected minute. A system that needs 108 attempts to produce 100 successful outcomes burns the failed eight on transcription, model tokens, generated speech, telephony, and infrastructure before they end. If users retry, another full set of costs starts. A low headline price cannot make up for poor completion rates.

Cost per successful outcome is the formula that matters: (Voice stack plus failure and retry overhead plus human handling plus evaluation and operations) divided by successful outcomes. Engineering, observability, and evaluation infrastructure show up here even though they do not appear on any per-minute line. The cheapest headline rate can be the most expensive real cost when failure rates are high.

What this article does not publish

No vendor recommendation is published here. Vendor pricing pages and effective per-minute rates change every quarter as new speech-to-speech models ship and committed-use tiers move. The durable answer is to re-check the linked pricing pages at the moment of procurement.

No latency guarantee is published here. Production latency depends on model choice, geography, and the connection between the agent and the telephony provider. Run a small pilot on your real traffic shape before the cost model decides the architecture.

A useful next action

Estimate the monthly connected minutes the project will use in the first 90 days. Pick the volume band the estimate lands in and read off the per-minute range from the table above. Multiply by 2x to 3x to land on a realistic loaded cost for any developer-platform option, then compare against a self-assembled cascaded stack budget that includes engineering time.

Run a small pilot (1,000 to 10,000 minutes) on two candidate stacks before committing to a contract. Track per-minute cost, completion rate, and latency p50 and p95 on the same dashboard, then re-budget the production plan from the pilot data.