Research-backed calculator guide

Voice AI Minimum Billing Increments: Model Short Calls, Rounding and Effective Rates

Learn how minimum commitments and billing increments change short voice-AI call costs, with formulas, a worked distribution model and invoice-reconciliation checks.

What the minimum actually applies to

“Voice AI cost per minute” can describe several different meters. A managed platform may charge for connected agent time. A speech-to-text endpoint may charge for audio successfully processed. A language model may charge input and output tokens. Text-to-speech may charge characters. Telephony may use its own minimum and increment. Vapi documents this layered structure explicitly: transcription is measured in audio minutes, the model in input and output tokens, and the voice in spoken characters ^[5].

That distinction prevents a common error: applying the telephony minimum to every layer, or treating a platform minute as if it were one minute of speech. If a provider sells a bundled agent minute, read what is included. Deepgram’s published Voice Agent table covers speech-to-text, a managed language model and text-to-speech at one connection-time rate, while its BYO options change the bundle ^[4]. Retell’s pricing page likewise publishes a per-minute range but describes separate voice infrastructure, text-to-speech, LLM, telephony and add-on components ^[6].

There are four questions to answer for every line item:

1. What starts the meter? It might be a request, an answered call, audio arrival or an open WebSocket.
2. What stops it? Check hang-up, transfer, timeout and post-call processing behavior.
3. What is the minimum? A provider can have no minimum, a minimum duration, a monthly minimum or all three.
4. What is the increment after the minimum? Common patterns include one second, 6 seconds, 15 seconds and 60 seconds, but the contract or product documentation controls.

Do not infer a provider’s rule from a competitor’s rule. A search result can show that one telephony service uses 60/60 billing while a speech API bills per second. It cannot establish how an uninspected voice-agent platform bills your call.

Four billing patterns worth modeling

Per-second billing. The provider charges the measured duration, usually rounded up to a whole second. This is relatively efficient for short calls, but it does not automatically mean that silence, hold time or an open session is free. Google states that successfully processed audio is counted and that multiple channels are billed separately ^[1]. Deepgram says its Voice Agent charge is tied to WebSocket connection time, so the connection boundary matters ^[4].

Fixed block billing. The first block may be larger than subsequent blocks. In a 30/6 schedule, a 10-second call is charged 30 seconds, while a 31-second call is charged 36 seconds. In a 60/60 schedule, a 61-second call is charged 120 seconds. This structure makes sub-minute workflows especially sensitive to the distribution of duration.

Per-minute rounding. Some services take a partial minute and round it to a full minute. Twilio’s help guidance gives a simple example: a one-minute-20-second call is billed as two minutes ^[3]. Its reconciliation documentation adds that usage records represent billed minutes per call, rounded according to the pricing model’s minimum and round quantity ^[2].

Layered or token-based billing. A voice platform may show a minute estimate while the final invoice moves with prompt length, tool definitions, conversation history, spoken characters or model turns. Vapi says its performance metric is an estimate rather than an exact quote and assumes component-specific usage, including a 50% cache hit rate when applicable ^[5]. Treat such figures as planning estimates and label every assumption.

Worked cost example

Assume a hypothetical bundled voice-agent rate of $0.08 per billed minute. This is an illustration, not a vendor quote. Ignore tax, carrier surcharges, model pass-throughs and monthly subscriptions so the effect of the billing increment is isolated. The raw duration is the connected call length before any rounding.

The table shows why the headline $0.08/min is incomplete. A 14-second call costs $0.0187 under exact per-second billing, but $0.08 under 60/60 billing. Its effective rate is therefore $0.08 divided by 14/60 minutes, or about $0.34 per raw minute. Under 30/6, the same call costs $0.04, an effective rate of about $0.17 per raw minute.

Now model a small call mix rather than one example:

• 600 calls last 14 seconds.
• 300 calls last 47 seconds.
• 100 calls last 61 seconds.
• The nominal rate is $0.08 per billed minute.

The raw total is (600 × 14) + (300 × 47) + (100 × 61) = 28,600 seconds, or 476.67 raw minutes. At exact per-second billing, the cost is 476.67 × $0.08 = $38.13.

Under 30/6 billing, the same calls consume (600 × 30) + (300 × 48) + (100 × 66) = 39,000 billed seconds, or 650 billed minutes. The modeled charge is $52.00, which is $13.87 higher than the per-second case. The effective cost is $52 / 476.67 = $0.109 per raw minute, not the advertised $0.08.

Under 60/60 billing, the distribution consumes 600 × 60 + 300 × 60 + 100 × 120 = 66,000 billed seconds, or 1,100 billed minutes. The modeled charge is $88.00, and the effective cost is about $0.185 per raw minute. The same nominal rate is now more than twice the exact per-second total because many calls finish before the first minute boundary.

This is the original analysis to carry into your own model: calculate the billed seconds for every call, sum those seconds, then divide total charge by raw connected minutes. Do not multiply the average raw duration by a rounded average. Rounding is nonlinear, so the order of operations matters.

How to build a reliable short-call model

Start with a usage export containing a call or request identifier, start and end timestamps, status, raw duration, billed duration, product or endpoint, and any channel count. Keep failed, busy and unanswered attempts separate because billing rules may exclude them. Twilio’s reconciliation guidance notes that not every call in logs is billed and that usage data represents billed minutes rather than raw log duration ^[2].

Next, create one row for every meter. For a bundled agent, that might be agent connection time and telephony. For an assembled stack, add speech input, model input and output tokens, synthesized characters, carrier time and storage. Mark a row as “included” only when the provider documentation or contract says so. A blank price field is not evidence of a free feature.

Then test edge cases at each boundary: one second, one second below the minimum, exactly at the minimum, one second above it, and one increment below the next boundary. These cases expose whether the schedule is “minimum plus subsequent increment,” “round every request,” or “round the total monthly usage.” If the documentation is unclear, ask the provider for an invoice example rather than selecting the most favorable interpretation.

Finally, reconcile the model after the first billing cycle. Compare your calculated billed units with the provider’s usage export, normalize time zones, and investigate mismatches caused by retries, transfers, multiple call legs, channels or sessions left open. An estimate is useful only when its assumptions can explain the invoice.

Try the Mistral and Groq API cost calculator with the same assumptions.

For broader comparisons, use the model pricing comparison to preserve each provider’s native billing unit rather than flattening every quote into one per-minute number. When your forecast and invoice diverge, the AI API invoice reconciliation calculator provides a useful adjacent workflow. If metering gaps are recurring, the AI observability ROI calculator can help frame the cost of adding usage traces and review time.

Assumptions, exclusions and limitations

This example assumes a single billed voice-agent meter, a constant USD rate and one connected call leg. It does not estimate a provider’s actual price, taxes, carrier rates, regional surcharges, concurrency fees, subscription minimums, voice licensing, storage, retries or LLM token usage. It also assumes that the raw duration is the correct denominator for the effective-rate calculation.

In production, “raw duration” needs a definition. A provider may start at answer, first audio, session creation or the first successfully processed request. Silence and hold time may remain billable when the session is open. A transfer can stop the AI meter while telephony continues, or it can create a second billed leg. Multiple audio channels can multiply processed duration; Google documents this explicitly for Speech-to-Text ^[1].

Prices and product rules change. The observations in this article are dated to 2026-11-02 where relevant, but the provider pages remain the source of truth for a future quote. Vendor examples are not interchangeable: Google’s one-second audio rule, Deepgram’s connection-time Voice Agent rule, Twilio’s call-minute rounding and Retell’s nearest-second agent rule apply to their named products, not to the voice AI market as a whole.

FAQ

What is a minimum billable duration for a voice agent?

It is the shortest duration a provider charges for a defined meter. A 30-second minimum means a 4-second eligible call can consume 30 billed seconds. It may apply per call, per request or per session. Confirm the meter’s start and stop events before using the term in a forecast.

How does voice AI per-minute rounding affect short calls?

Rounding increases billed seconds whenever the raw duration falls below a billing boundary. The effect is largest when calls are shorter than one block. In the worked example, 14-second calls consume 30 seconds under 30/6 and 60 seconds under 60/60, so the effective raw-minute cost is several times the nominal per-minute rate.

Does per-second billing mean there is no minimum charge?

No. Per-second may mean that a measured duration is rounded to one second, but a provider can still impose a minimum request duration, monthly commitment, platform fee or minimum invoice. Read the pricing unit and the commercial minimum separately. Deepgram and Google publish per-second language for specific products, while that does not establish the rules for a different endpoint ^{[1] [4]}.

Are silence, hold time and tool latency billed?

They can be if the billed meter is connected or session time. A speech-recognition endpoint may charge processed audio while a managed agent may charge an open connection. The safest approach is to compare start and stop timestamps in usage records and ask whether waiting, transfer setup and tool calls remain inside the billable span.

Should I use average call length to forecast spend?

Use the full distribution whenever minimums or block rounding apply. Multiplying an average by a rounded unit can hide the number of calls just below a boundary. If you only have summary data, calculate a low, central and high case using plausible short-call shares, then replace those assumptions with observed durations after a pilot.

How do I compare a bundled voice-agent minute with a speech API rate?

List the included layers first. A bundled minute may include orchestration, speech recognition and speech synthesis, while a speech API rate may cover only audio processing. Keep tokens, characters, telephony, concurrency and subscriptions as separate rows. Compare the total for the same call mix and the same success definition, not one sticker rate against another.