AI voice agents feel expensive because every call stacks multiple meters: speech-to-text, the language model, text-to-speech, telephony, observability, and sometimes a human escalation. The LLM layer is not the whole bill, but it is the easiest layer to overpay for because long conversations multiply tokens turn by turn.
This guide breaks down the model cost layer for AI voice agents in 2026. We will calculate cost per call, cost per 10,000 conversations, and monthly budgets for inbound support, appointment booking, qualification calls, and internal phone automation. The numbers below use real per-token model pricing from AI Cost Check’s model data, including GPT-5 mini, Gemini 2.5 Flash, DeepSeek V4 Flash, Claude Sonnet 4.6, and GPT-5.2.
The key recommendation: run routine voice turns on low-cost fast models, then route only complex or high-value calls to premium escalation models. For most support and booking agents, that reduces the LLM layer from dollars per hundred calls to cents per hundred calls without changing your speech stack.
💡 Key Takeaway: For a 6-minute support call using about 10,800 input tokens and 1,800 output tokens, the LLM layer can cost $0.0020 on DeepSeek V4 Flash, $0.0077 on GPT-5 mini, or $0.0594 on Claude Sonnet 4.6.
What counts as the LLM cost layer in a voice agent?
A production AI voice agent usually has five billable layers:
- Telephony — phone numbers, SIP, inbound/outbound minutes, recording, call transfer.
- Speech-to-text — transcribing the caller’s audio into text.
- LLM reasoning and response generation — deciding what to say, using tools, following policy, and producing the next reply.
- Text-to-speech — turning the response text into audio.
- Platform overhead — orchestration, logging, analytics, retries, guardrails, and human handoff.
This article focuses on layer 3: the LLM. That is the cost you can compare across models with token pricing. Speech-to-text and text-to-speech sit outside the model budget unless your vendor bundles them into a single realtime API price. When a vendor quotes “voice agent cost per minute,” ask for the split between transcription, synthesis, telephony, and LLM tokens.
The LLM layer is priced by tokens. Input tokens include the system prompt, conversation history, retrieved policy snippets, tool results, and the caller’s latest transcript. Output tokens are the model’s generated responses, tool calls, structured JSON, and internal routing decisions if billed as output.
A short booking call can use only a few thousand tokens. A troubleshooting call can use tens of thousands because every new turn may include accumulated context. A voice agent that repeats the full policy document on every turn will spend far more than an agent that retrieves only the relevant paragraph.
The voice-agent token formula
Use this formula for each call:
LLM cost per call = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)
For example, GPT-5 mini costs $0.25 per 1M input tokens and $2 per 1M output tokens. A call with 10,800 input tokens and 1,800 output tokens costs:
- Input: 10,800 ÷ 1,000,000 × $0.25 = $0.0027
- Output: 1,800 ÷ 1,000,000 × $2 = $0.0036
- Total: $0.0063 per call
The same call on Claude Sonnet 4.6, at $3 input and $15 output per 1M tokens, costs:
- Input: 10,800 ÷ 1,000,000 × $3 = $0.0324
- Output: 1,800 ÷ 1,000,000 × $15 = $0.0270
- Total: $0.0594 per call
That gap becomes material at call-center volume.
Baseline token assumptions for common voice calls
Voice agents are turn-based systems. A “turn” usually means the caller says something, the transcript is sent to the model, the model thinks, may call a tool, and then replies. The number of turns matters more than call duration alone.
Here are practical token budgets for LLM cost modeling:
| Call type | Typical duration | Turns | Input tokens per call | Output tokens per call | Use case |
|---|---|---|---|---|---|
| Appointment booking | 3 minutes | 6 | 4,500 | 900 | Scheduling, rescheduling, confirmations |
| Lead qualification | 5 minutes | 10 | 8,000 | 1,500 | Sales intake, qualification, routing |
| Inbound support | 6 minutes | 12 | 10,800 | 1,800 | FAQs, account help, troubleshooting |
| Internal phone automation | 4 minutes | 8 | 6,000 | 1,200 | IT helpdesk, HR, facilities, internal ops |
| Complex escalation | 10 minutes | 18 | 25,000 | 4,000 | Refund disputes, medical/financial admin, multi-step workflows |
These budgets assume compact prompts, limited retrieved context, and short spoken responses. If your agent injects a 20,000-token policy manual into every turn, costs will be dramatically higher. Use retrieval, summaries, and state variables rather than dumping the entire call history and knowledge base repeatedly.
⚠️ Warning: The most common voice-agent cost mistake is counting only the caller transcript. The model also receives system prompts, conversation history, retrieved knowledge, tool outputs, and hidden orchestration text. Those tokens are usually larger than the caller’s words.
Model pricing for AI voice agent LLMs
The cheapest model is not automatically the best voice model. Voice agents need low latency, strong instruction following, stable tool use, and concise responses. Still, the pricing spread is wide enough that model selection should be a first-order architecture decision.
The table below compares representative models for real-time support and escalation. Pricing is per 1 million tokens.
| Model | Provider | Input price | Output price | Context window | Best role in voice stack |
|---|---|---|---|---|---|
| DeepSeek V4 Flash | DeepSeek | $0.14 | $0.28 | 1,000,000 | Lowest-cost routine calls and high-volume triage |
| Gemini 2.0 Flash-Lite | $0.075 | $0.30 | 1,000,000 | Very cheap classification, routing, short responses | |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | 1,000,000 | Budget general-purpose support turns | |
| GPT-5 nano | OpenAI | $0.05 | $0.40 | 128,000 | Ultra-low-cost intent classification and simple scripts |
| GPT-5 mini | OpenAI | $0.25 | $2.00 | 500,000 | Balanced mainstream support and booking |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1,000,000 | Fast support with stronger general reasoning | |
| Mistral Small 4 | Mistral AI | $0.15 | $0.60 | 128,000 | Cost-efficient scripted support |
| GPT-5.2 | OpenAI | $1.75 | $14.00 | 1,000,000 | Premium escalation and complex workflows |
| Claude Sonnet 4.6 | Anthropic | $3.00 | $15.00 | 1,000,000 | Premium support, policy-heavy calls, high-stakes escalation |
| Claude Opus 4.7 | Anthropic | $5.00 | $25.00 | 1,000,000 | Highest-cost expert escalation layer |
For high-volume voice agents, DeepSeek V4 Flash, Gemini 2.0 Flash-Lite, Gemini 2.5 Flash-Lite, and GPT-5 nano are the cost leaders. GPT-5 mini is a good default when you want a stronger balance between cost, tooling, and general capability. Claude Sonnet 4.6 and GPT-5.2 should be escalation models, not the default for every routine call.
Cost per call by model
Using the baseline 6-minute inbound support call at 10,800 input tokens and 1,800 output tokens, here is the LLM cost by model.
| Model | Input cost | Output cost | Total per call | Cost per 10,000 calls |
|---|---|---|---|---|
| DeepSeek V4 Flash | $0.001512 | $0.000504 | $0.002016 | $20.16 |
| Gemini 2.0 Flash-Lite | $0.000810 | $0.000540 | $0.001350 | $13.50 |
| Gemini 2.5 Flash-Lite | $0.001080 | $0.000720 | $0.001800 | $18.00 |
| GPT-5 nano | $0.000540 | $0.000720 | $0.001260 | $12.60 |
| GPT-5 mini | $0.002700 | $0.003600 | $0.006300 | $63.00 |
| Gemini 2.5 Flash | $0.003240 | $0.004500 | $0.007740 | $77.40 |
| Mistral Small 4 | $0.001620 | $0.001080 | $0.002700 | $27.00 |
| GPT-5.2 | $0.018900 | $0.025200 | $0.044100 | $441.00 |
| Claude Sonnet 4.6 | $0.032400 | $0.027000 | $0.059400 | $594.00 |
| Claude Opus 4.7 | $0.054000 | $0.045000 | $0.099000 | $990.00 |
The LLM layer for 10,000 support calls ranges from $12.60 on GPT-5 nano to $990.00 on Claude Opus 4.7 under this token budget. That is a 78.6x spread for the same call volume.
[stat] 78.6x The cost difference between GPT-5 nano and Claude Opus 4.7 for 10,000 baseline inbound support calls
The right architecture is not “use the cheapest model everywhere.” The right architecture is to route the first-line conversation through a cheap model, detect risk and complexity, and escalate selectively.
Scenario 1: Appointment booking voice agent
Appointment booking is the easiest voice-agent workload to make cheap. The task is structured: collect name, time preference, service type, location, and confirmation. Most calls are short, and the agent can rely on tool calls to the calendar system rather than long reasoning.
Assumption per call:
- 4,500 input tokens
- 900 output tokens
- 3-minute average call
- 6 model turns
- Calendar lookup and booking tool call included in the prompt budget
| Model | Cost per call | Cost per 10,000 calls | Monthly cost at 30,000 calls |
|---|---|---|---|
| GPT-5 nano | $0.000585 | $5.85 | $17.55 |
| DeepSeek V4 Flash | $0.000882 | $8.82 | $26.46 |
| Gemini 2.5 Flash-Lite | $0.000810 | $8.10 | $24.30 |
| GPT-5 mini | $0.002925 | $29.25 | $87.75 |
| Claude Sonnet 4.6 | $0.027000 | $270.00 | $810.00 |
Recommendation: use GPT-5 nano, Gemini 2.5 Flash-Lite, or DeepSeek V4 Flash for the default booking flow. Use GPT-5 mini only if your booking logic includes nuanced policy interpretation, multilingual complexity, or many edge cases. Premium models are wasteful for routine appointment booking.
A booking agent handling 30,000 calls/month can keep the LLM layer under $30/month with the cheapest models. Even GPT-5 mini is only $87.75/month in this scenario. Your telephony and speech costs will likely dominate the total vendor bill.
📊 Quick Math: At 30,000 appointment calls/month, Claude Sonnet 4.6 costs $810/month for the LLM layer, while Gemini 2.5 Flash-Lite costs $24.30/month. The savings are $785.70/month before touching speech or telephony.
Scenario 2: Inbound customer support
Inbound support is the most common AI voice-agent deployment and the easiest to under-budget. Calls are longer than booking flows, users ask follow-up questions, and the model often needs retrieved knowledge or account-specific tool results.
Assumption per call:
- 10,800 input tokens
- 1,800 output tokens
- 6-minute average call
- 12 model turns
- Moderate retrieval context and one or two tool calls
| Model | Cost per call | Cost per 10,000 calls | Monthly cost at 100,000 calls |
|---|---|---|---|
| GPT-5 nano | $0.001260 | $12.60 | $126.00 |
| DeepSeek V4 Flash | $0.002016 | $20.16 | $201.60 |
| Gemini 2.5 Flash-Lite | $0.001800 | $18.00 | $180.00 |
| GPT-5 mini | $0.006300 | $63.00 | $630.00 |
| Gemini 2.5 Flash | $0.007740 | $77.40 | $774.00 |
| Claude Sonnet 4.6 | $0.059400 | $594.00 | $5,940.00 |
| Claude Opus 4.7 | $0.099000 | $990.00 | $9,900.00 |
Recommendation: default to DeepSeek V4 Flash, Gemini 2.5 Flash-Lite, or GPT-5 mini. Use Claude Sonnet 4.6 or GPT-5.2 only for escalated cases such as refund disputes, compliance-sensitive answers, or complex troubleshooting. If you are evaluating premium model tradeoffs, start with GPT-5 vs Claude Sonnet 4.5 and GPT-5 vs DeepSeek V3.2 to understand the cost curve across tiers.
A high-volume support operation with 100,000 calls/month can keep routine LLM spend under $1,000/month with GPT-5 mini or below $250/month with budget models. Running every call on Claude Sonnet 4.6 raises the LLM layer to $5,940/month for the same token budget.
Scenario 3: Lead qualification and sales intake
Lead qualification calls need better conversational judgment than appointment booking. The agent must ask discovery questions, identify budget and urgency, handle objections, and decide whether to book a meeting, create a CRM record, or disqualify the lead.
Assumption per call:
- 8,000 input tokens
- 1,500 output tokens
- 5-minute average call
- 10 model turns
- CRM lookup and structured summary included
| Model | Cost per call | Cost per 10,000 calls | Monthly cost at 50,000 calls |
|---|---|---|---|
| DeepSeek V4 Flash | $0.001540 | $15.40 | $77.00 |
| Gemini 2.5 Flash-Lite | $0.001400 | $14.00 | $70.00 |
| GPT-5 mini | $0.005000 | $50.00 | $250.00 |
| Gemini 2.5 Flash | $0.006150 | $61.50 | $307.50 |
| GPT-5.2 | $0.035000 | $350.00 | $1,750.00 |
| Claude Sonnet 4.6 | $0.046500 | $465.00 | $2,325.00 |
Recommendation: use GPT-5 mini or Gemini 2.5 Flash when conversion quality matters. Use Gemini 2.5 Flash-Lite or DeepSeek V4 Flash when the qualification script is rigid and your CRM workflow drives most decisions. Escalate only high-intent or high-contract-value leads to GPT-5.2 or Claude Sonnet 4.6.
This is where model routing creates measurable business value. If 10% of calls are high-value and need a premium model, you can run 90% on GPT-5 mini and 10% on GPT-5.2. At 50,000 calls/month, using the lead qualification token budget:
- 45,000 GPT-5 mini calls × $0.005 = $225
- 5,000 GPT-5.2 calls × $0.035 = $175
- Total blended LLM cost = $400/month
Running all 50,000 calls on GPT-5.2 would cost $1,750/month. Routing saves $1,350/month, a 77% reduction, while preserving premium reasoning for the calls that justify it.
Scenario 4: Internal phone automation
Internal phone automation includes IT helpdesk password flows, HR policy questions, facilities requests, store operations, and field-team updates. These calls tend to be lower risk than customer-facing support but may require authentication and accurate tool use.
Assumption per call:
- 6,000 input tokens
- 1,200 output tokens
- 4-minute average call
- 8 model turns
- One authentication or ticketing tool call
| Model | Cost per call | Cost per 10,000 calls | Monthly cost at 20,000 calls |
|---|---|---|---|
| GPT-5 nano | $0.000780 | $7.80 | $15.60 |
| DeepSeek V4 Flash | $0.001176 | $11.76 | $23.52 |
| Mistral Small 4 | $0.001620 | $16.20 | $32.40 |
| GPT-5 mini | $0.003900 | $39.00 | $78.00 |
| Claude Sonnet 4.6 | $0.036000 | $360.00 | $720.00 |
Recommendation: use GPT-5 nano, DeepSeek V4 Flash, or Mistral Small 4 for internal phone automation. Internal use cases usually benefit more from reliable integrations and audit logs than premium model reasoning. Save premium models for exception handling, legal policy interpretation, or sensitive employee relations workflows.
✅ TL;DR: Routine voice agents are cheap at the LLM layer. The expensive pattern is sending every call, every turn, and every retrieved document to a premium model when a fast budget model can handle the first-line flow.
How latency affects spend
Latency and cost are connected because poor latency increases turn count. If the agent pauses too long, callers repeat themselves, interrupt, ask “are you still there?”, or restart the question. Those extra turns add input and output tokens.
A low-latency model with slightly higher token price can be cheaper than a slower model if it reduces conversation length. For voice agents, optimize for completed calls per dollar, not only token price.
Use these latency-driven budgeting rules:
- Keep spoken replies short: 25-60 words for routine answers.
- Use tool calls instead of verbal reasoning: “Let me check that” followed by a concise answer.
- Summarize old context after 6-8 turns instead of resending full transcripts.
- Retrieve only the top 1-3 relevant knowledge chunks.
- Route long-tail troubleshooting to a premium model after the cheap model detects complexity.
A voice call with 12 turns at 10,800 input / 1,800 output tokens costs $0.0063 on GPT-5 mini. If poor latency and verbose replies increase it to 18 turns and 18,000 input / 3,200 output tokens, GPT-5 mini rises to:
- Input: 18,000 × $0.25 / 1M = $0.0045
- Output: 3,200 × $2 / 1M = $0.0064
- Total: $0.0109 per call
That is a 73% increase without changing the model. At 100,000 calls/month, the difference is $460/month on GPT-5 mini. On Claude Sonnet 4.6, the same token growth raises cost from $5,940/month to $10,200/month, adding $4,260/month.
Why output tokens matter in voice agents
Many teams focus on input price because the context window grows over the call. In voice agents, output tokens deserve equal attention because premium models often charge much more for output.
For example:
- DeepSeek V4 Flash: $0.14 input / $0.28 output
- GPT-5 mini: $0.25 input / $2 output
- Claude Sonnet 4.6: $3 input / $15 output
- Claude Opus 4.7: $5 input / $25 output
If your agent gives long explanations, reads policy text aloud, or repeats disclaimers on every turn, output costs rise quickly. Shorter responses also improve user experience. Voice is slower than text, so a concise spoken answer is both cheaper and better.
A strong default system instruction is: “Answer in one or two short sentences unless the caller asks for detail.” That simple rule can cut output tokens by 30-60% in support calls.
Recommended model strategy for 2026 voice agents
Use a tiered model stack instead of a single model.
Tier 1: Classification and routing
Use GPT-5 nano, Gemini 2.0 Flash-Lite, or Gemini 2.5 Flash-Lite for:
- Intent detection
- Language detection
- Sentiment or escalation scoring
- Call-type classification
- Simple policy routing
- Post-call tagging
These tasks are short and structured. Paying premium-model rates for them is unnecessary.
Tier 2: Default real-time conversation
Use DeepSeek V4 Flash, Gemini 2.5 Flash-Lite, Mistral Small 4, or GPT-5 mini for:
- Appointment booking
- Tier-1 support
- FAQ handling
- Internal requests
- Lead intake
- Order status and account updates
For many production teams, GPT-5 mini is the safe default because it remains inexpensive while offering more headroom than nano-tier models. If maximum cost efficiency is the priority, DeepSeek V4 Flash and Gemini Flash-Lite tiers produce extremely low per-call LLM costs.
Tier 3: Premium escalation
Use GPT-5.2, Claude Sonnet 4.6, or Claude Opus 4.7 for:
- Refund disputes
- Multi-step troubleshooting
- Regulated or policy-heavy answers
- High-value sales opportunities
- Supervisor-style review
- Complex summarization after the call
Premium models should see a small percentage of traffic. A strong routing target is 5-15% of calls escalated to a premium model. If more than 25% of calls need premium escalation, your first-line flow, retrieval, or product documentation needs improvement.
For broader model tradeoffs, compare GPT-5 vs Claude Opus 4.6, GPT-5 vs Gemini 3 Pro, and Claude Opus 4.6 vs DeepSeek V3.2.
What is outside this budget?
The calculations above cover only LLM token cost. A complete voice-agent budget also includes:
- Speech-to-text: audio transcription or realtime speech recognition.
- Text-to-speech: generated voice audio.
- Phone network: PSTN, SIP, carrier, number rental, inbound/outbound minutes.
- Voice platform: call orchestration, barge-in, interruption handling, recording, analytics.
- Knowledge retrieval: vector database, search API, embedding model, cache.
- Tool execution: CRM, scheduling, payments, ticketing, identity checks.
- Monitoring and QA: call review, redaction, compliance logging, evaluation runs.
This separation matters because a vendor may quote $0.05 to $0.25 per minute for a bundled voice agent platform while the LLM layer is only $0.001 to $0.01 per call on budget models. If you are negotiating a vendor contract, ask for token usage export and model-level billing. Without it, you cannot know whether you are paying for speech, telephony, orchestration, or an overpriced default LLM.
Use AI Cost Check to model the LLM layer separately before comparing platform quotes. That gives you a baseline for what the language model should cost at your call volume.
Cost optimization checklist
Use this checklist before launching a production voice agent:
- Cap response length. Keep routine replies under 60 words.
- Summarize history. Replace old turns with a compact state summary after several exchanges.
- Use retrieval sparingly. Send only relevant knowledge chunks, not full documents.
- Route by risk. Keep routine calls on budget models and escalate complex calls.
- Cache static policy answers. Reuse common responses when the answer does not change.
- Separate classification from conversation. Use nano or Flash-Lite models for routing.
- Measure tokens per call type. Track booking, support, sales, and escalation separately.
- Set model budgets. Alert when average input or output tokens per call rise above target.
- Audit tool outputs. Long JSON payloads can become hidden input-token bloat.
- Test at 10,000-call scale. Small pilots hide cost outliers.
The most important metric is not cost per token. It is resolved calls per dollar. A model that costs 3x more but resolves 20% more calls may be justified for high-value support. A model that costs 30x more and produces the same booking completion rate is not.
Frequently asked questions
How much does an AI voice agent cost per call?
The LLM layer usually costs $0.001 to $0.01 per routine call on budget and mid-tier models, using the token assumptions in this guide. A 6-minute support call costs about $0.00126 on GPT-5 nano, $0.002016 on DeepSeek V4 Flash, $0.0063 on GPT-5 mini, and $0.0594 on Claude Sonnet 4.6.
How much does the LLM layer cost for 10,000 AI voice conversations?
For 10,000 baseline inbound support conversations, expect $12.60 on GPT-5 nano, $20.16 on DeepSeek V4 Flash, $63.00 on GPT-5 mini, $441.00 on GPT-5.2, and $594.00 on Claude Sonnet 4.6. Use the AI Cost Check calculator to adjust the input and output token assumptions for your own call flow.
What is the cheapest model for AI voice agents?
The cheapest models for the LLM layer include GPT-5 nano at $0.05 input / $0.40 output, Gemini 2.0 Flash-Lite at $0.075 input / $0.30 output, and DeepSeek V4 Flash at $0.14 input / $0.28 output per 1M tokens. For production support, DeepSeek V4 Flash, Gemini Flash-Lite, and GPT-5 mini are strong default choices depending on your latency, tooling, and quality requirements.
Are speech-to-text and text-to-speech included in these calculations?
No. These calculations cover only the LLM token layer. Speech-to-text, text-to-speech, phone minutes, call recording, orchestration, and vendor platform fees are separate budget lines unless your provider bundles everything into one realtime voice price.
When should I use a premium model for voice support?
Use a premium model for 5-15% of calls: high-value sales opportunities, refund disputes, regulated answers, complex troubleshooting, or supervisor-style escalation. Run routine booking, FAQs, order status, and internal requests on cheaper real-time models, then escalate when the model detects risk or complexity.
Calculate your voice-agent model budget
Before signing a voice-agent platform contract, estimate the LLM layer independently:
- Pick your call type: booking, support, qualification, internal automation, or escalation.
- Estimate input and output tokens per call.
- Multiply by your monthly conversation volume.
- Compare budget, mid-tier, and premium models.
- Add speech-to-text, text-to-speech, telephony, and platform fees separately.
Use AI Cost Check to calculate your own per-call and monthly costs across current models. For deeper model comparisons, review GPT-5 vs DeepSeek V3.2, GPT-5 vs GPT-5 mini, and Claude Opus 4.6 vs Gemini 3 Pro.