Research-backed calculator guide

Voice Agent Peak Concurrency Cost: Size Channels Separately From Monthly Minutes

Calculate voice agent peak concurrency cost without confusing simultaneous-call capacity with monthly minutes, platform rates, carrier charges, and burst headroom.

What peak concurrency means in a voice-agent budget

Concurrency is the count of active calls at the same moment. It is a capacity unit. Monthly minutes are a usage unit. One tells you how many live sessions the system must hold; the other tells you how long those sessions run across a billing period.

A monthly average hides the peak. Suppose 12,000 minutes are spread evenly across a month. That could be a steady 8 active calls, or it could be quiet overnight with 70 callers during a campaign hour. The same minute total can require very different channel capacity.

Capacity also exists at several layers. Your voice-agent platform may limit active sessions per workspace or project. Your carrier or SIP trunk may limit channels or calls per second. A speech provider may apply a project concurrency limit. The practical ceiling is the tightest limit along the call path. Deepgram describes voice-agent pricing as connection-time billing and notes that concurrency limits differ by plan and region; its pricing page separately shows per-minute rates for bundled and bring-your-own components. ^[2]

Do not multiply a per-minute price by “peak concurrency.” That would charge capacity as though every channel were continuously used. Instead, use peak concurrency for the capacity line and actual billable minutes for the usage line.

The calculation: minutes and channels are separate

Start with a measured call-detail export when possible. Useful fields include call start and end timestamps, direction, outcome, transfer status, and the platform’s billable duration. Aggregate it into:

• monthly connected calls and billable minutes;
• busy-hour calls for ordinary high-volume periods;
• worst 15-minute or 30-minute arrival count during a plausible surge;
• average and high-percentile duration for those intervals;
• calls per second at the beginning of a campaign or appointment window; and
• carrier, platform, and model limits.

For a first estimate, use this formula:

base concurrency = busy-hour calls × average duration (minutes) ÷ 60
required channels = ceiling(base concurrency × peak headroom factor)
capacity overage = max(0, required channels − included channels) × extra-channel fee

The headroom factor is an assumption, not a law. Use a documented factor such as 1.25 or 1.35 for an early budget, then replace it with interval-level load-test results. It covers clustering inside the hour, duration variation, transfers that keep a call open, and a small amount of forecast error. It does not prove that a carrier will accept every burst. Validate calls per second separately.

The cost model is:

monthly cost = platform usage + carrier usage + capacity + numbers
               + enabled add-ons + transfer/recording/storage costs

Avoid double counting. A bundled voice-agent rate may include speech recognition, a managed language model, and text-to-speech. A platform such as Vapi instead presents hosting, model pass-through, and transport as separate lines. ^[5] If you use a bundled rate, do not add the same STT or TTS component again. If you use a modular stack, record each component’s unit and rounding rule.

Worked cost example

Assume an English-US inbound support line with these illustrative inputs. They are deliberately explicit so you can replace them with your own logs. The platform and carrier figures below are published examples checked on 2026-11-02, not a quote or forecast of your bill.

• 2,400 connected calls per month;
• 4.0 average billable minutes per call, or 9,600 monthly minutes;
• 180 calls in the busiest ordinary hour;
• 4.5 minutes average duration during that hour;
• 1.35 peak headroom factor;
• 20 included concurrent calls; and
• $8 per extra concurrent call per month, using Retell’s published capacity line as the example. ^[3]

The concurrency math is 180 × 4.5 ÷ 60 = 13.5 base concurrent calls. Applying 1.35 gives 18.225, which rounds up to 19 required channels. Because 19 is below the illustrative 20 included channels, the capacity charge is $0. The monthly usage line still applies: 9,600 minutes × $0.11 per minute = $1,056 when using the Retell page’s displayed estimate as the blended platform example. The resulting illustrative subtotal is $1,056 before any phone-number, add-on, tax, or nonstandard transfer charges.

Now model a launch campaign that produces 300 calls in the busiest hour at the same 4.5-minute average. The base concurrency becomes 22.5. With 1.35 headroom, required capacity is 30.375, rounded up to 31 channels. That is 11 channels above the included 20. At the published $8 example, capacity adds 11 × $8 = $88 per month. If the campaign also increases monthly minutes to 12,000, the blended usage example becomes 12,000 × $0.11 = $1,320, for an illustrative subtotal of $1,408 before other charges.

The table exposes the key budgeting lesson: the surge adds $264 of usage in this example because minutes rise, but only $88 of recurring capacity because channels are priced independently. A different vendor may include fewer channels, charge a different line rate, or use a burst surcharge instead. Keep those policies in separate columns.

Price a modular stack without double counting

A bundled platform is not the only architecture. Deepgram’s published Voice Agent Standard rate is $0.075 per minute on pay-as-you-go and covers speech-to-text, a managed language model, and text-to-speech; its page also lists lower rates for bring-your-own components. ^[2] For the same 9,600 minutes, the bundled usage arithmetic is 9,600 × $0.075 = $720.

Telephony is separate in a modular estimate. Twilio’s US page, marked current as of August 2026, lists local inbound calls at $0.0085 per minute and local numbers at $1.15 per month. ^[4] If every illustrative minute is a local inbound minute, carrier usage is 9,600 × $0.0085 = $81.60, plus $1.15 for one local number. That produces $802.75 before taxes, add-ons, transfers, or capacity fees: $720 + $81.60 + $1.15.

This is not a claim that Deepgram plus Twilio is cheaper for your deployment. The assumptions do not include your prompts, integrations, call recording, failed-call behavior, outbound traffic, or any platform concurrency purchase. The point is methodological: a bundled minute can be compared with a modular stack only after every included and excluded component is written down.

What changes peak cost most often?

Call duration distribution. Replacing a four-minute average with a nine-minute average more than doubles the concurrency implied by the same arrival rate. Use interval-level averages and a high-duration scenario instead of one annual average.

Arrival clustering. Calls that arrive at the top of an hour can hit calls-per-second limits before they exhaust steady-state channels. Treat CPS as its own constraint.

Transfers and hold time. A transfer may keep a telephony leg open after AI billing stops, or it may keep platform billing active depending on the vendor. Retell states on its pricing page that its AI fee stops after transfer while telephony continues, and that silence or hold time is billed during an active AI call. ^[3] Confirm the rule for your stack.

Bundling. Deepgram and Vapi illustrate opposite modeling risks: one can bundle several voice-agent layers into a minute rate, while the other exposes hosting, models, and transport separately. ^{[2] [5]} Your spreadsheet should have an “included in rate” column.

Overflow policy. A low monthly capacity fee is not useful if excess calls are rejected, queued too long, or sent to an unstaffed fallback. Cost the operational consequence of a full pool, not only the extra line item.

A practical capacity-budget workflow

First, export 30–90 days of call records and keep the timestamp granularity. Second, calculate monthly minutes and 15-minute, 30-minute, and hourly peaks. Third, create normal, expected, and seasonal scenarios. Fourth, calculate required channels with an explicit headroom factor. Fifth, map platform, provider, carrier, CPS, and fallback limits. Sixth, apply the actual included capacity and overage policy. Finally, reconcile the estimate to the invoice after launch. The same discipline used in an AI API invoice reconciliation calculator helps you distinguish rate variance from call-shape variance.

Use the Model pricing comparison route when changing the language model or speech layer. For ongoing alerts, retention, and trace economics, see the AI observability ROI calculator. The Mistral and Groq API cost calculator at /mistral-groq-cost-calculator is the assigned calculator link for this guide, although voice-agent estimates still require explicit minutes, carrier, and channel inputs.

FAQ

Is peak concurrency the same as monthly voice minutes?

No. Concurrency is simultaneous active calls. Monthly minutes are accumulated usage over a billing period. A workload can have low monthly minutes but a sharp campaign peak, or high minutes spread evenly with modest concurrency. Budget both dimensions.

How do I calculate voice AI channels from calls per hour?

Multiply calls in the busy hour by average call duration in minutes, divide by 60, multiply by a documented headroom factor, and round up. For 180 calls averaging 4.5 minutes with 1.35 headroom, the result is ceiling(180 × 4.5 ÷ 60 × 1.35) = 19 channels. Validate with shorter intervals if arrivals cluster.

Should I size to the average or the worst peak?

Use separate normal and surge scenarios. Size normal capacity to an evidence-based busy-hour percentile, then decide whether a seasonal peak uses paid channels, temporary burst capacity, queueing, or human overflow. Permanently buying for an exceptional event can waste money, while ignoring it can cause rejected calls.

Does a concurrency fee replace the per-minute fee?

Usually not. A concurrency fee normally purchases simultaneous capacity. You still pay for connected or billable minutes unless the vendor explicitly bundles them. Read whether a plan includes channels, minutes, both, or neither. Vapi’s pricing page, for example, presents hosting and model/transport charges alongside included concurrency and extra-line pricing. ^[5]

What should be included in a voice agent peak capacity budget?

Include platform usage, model or speech components not bundled into that usage rate, carrier minutes, phone numbers, extra channels, CPS capacity, recording and transcription, transfers, add-ons, taxes, and a defined overflow path. Also record exclusions such as implementation labor, human-agent wages, refunds, and failed calls.

How often should the estimate be recalculated?

Recalculate after a routing, prompt, carrier, model, or telephony change, and at least monthly while traffic is ramping. Review it before campaigns and seasonal events. Reconcile forecast minutes and peak channels to observed logs so the headroom factor becomes evidence rather than a permanent guess.