Twilio Media Streams + Bring-Your-Own-LLM: Cost Breakdown 2026

Twilio's $0.004/min Media Streams plus inbound voice plus your own LLM bridge can land under $0.05 per minute total. Here is what to budget and where the hidden costs hide.

The cost problem

flowchart LR
  Twilio["Twilio Media Streams"] -- "WS · μlaw 8kHz" --> Bridge["FastAPI Bridge :8084"]
  Bridge -- "PCM16 24kHz" --> OAI["OpenAI Realtime"]
  OAI --> Bridge
  Bridge --> Twilio
  Bridge --> Logs[(structured logs · OTel)]

Plenty of teams build voice agents on Twilio Programmable Voice + Media Streams and bring their own LLM (OpenAI, Anthropic, or self-hosted). The pitch is full control and predictable telephony cost. The reality is that "Twilio cost" is multiple line items stacked, and the LLM is usually the biggest one.

If you do not break out every line item, you will under-budget by 30–60% and find out at month-end.

How Twilio prices it

Twilio's pricing has roughly five layers for an inbound voice AI agent:

• Phone number (US local): $1.15/month per number
• Inbound call to that number: $0.0085/min in the US
• Outbound dial (if you call out): $0.014/min in the US
• Media Streams: $0.004/min on top of the call
• Toll-free numbers: $2/month + $0.022/min inbound

Those telephony costs apply regardless of the LLM. They are the "rails" cost. Then on top:

• STT (Deepgram Nova-3): $0.0048/min, or you let your LLM do speech-in directly
• LLM compute: depends on provider
• TTS (Aura-2 or ElevenLabs): $0.030 per 1k chars or $0.05–$0.10 per 1k chars

Honest math

Profile A — Inbound 5-minute call, GPT-4o-mini brain, Deepgram STT, Aura-2 TTS:

• Phone number amortized: ~$0.001/min if you handle 1k min/mo per number
• Inbound: 5 × $0.0085 = $0.0425
• Media Streams: 5 × $0.004 = $0.020
• STT: 5 × $0.0048 = $0.024
• LLM (GPT-4o-mini cached): ~$0.024
• TTS Aura-2 (2 min agent speech): $0.045
• Total: ~$0.156/call → $0.031/min

Profile B — Inbound 5-min call, gpt-realtime end-to-end via Twilio bridge:

• Phone number: ~$0.001/min
• Inbound: $0.042
• Media Streams: $0.020
• gpt-realtime cached: ~$0.28
• Total: ~$0.343 → $0.069/min

Profile C — Outbound 3-minute qualification, GPT-4o-mini + Aura-2:

• Phone number amortized: ~$0.001/min
• Outbound: 3 × $0.014 = $0.042
• Media Streams: $0.012
• STT + LLM + TTS: ~$0.045
• Total: $0.10/call → $0.033/min

The takeaway: Twilio + cascaded brings you to ~$0.03/min all-in. Twilio + end-to-end Realtime brings you to ~$0.07/min all-in. Both are SMB-margin friendly.

Hidden costs to watch

1. Recording storage — $0.0025/min stored (free for 10k min/mo on Voice).
2. Conversational Intelligence if you turn on Twilio's bundled features — adds $0.01–$0.03/min.
3. International inbound — can be 5–20× US rates; check origin country.
4. Number warmup — A2P 10DLC compliance fees if you also send SMS off the same brand.
5. Egress if you stream Media Streams to an EU box from a US Twilio account — small but real.

How CallSphere optimizes

CallSphere builds Twilio + BYO-LLM bridges across the 6 verticals — the Salon GlamBook (4 agents, GB-### booking refs), the Sales product, and the OneRoof Real Estate suite all use this pattern. The Healthcare Voice Agent uses a different telephony provider for HIPAA reasons but the bridge architecture is the same.

We run a tight cost ledger: every call gets logged to Postgres with line items for telephony, STT, LLM, TTS, and Media Streams minutes. The 90+ tools across 115+ DB tables give us per-tenant per-vertical attribution. In April 2026 our blended Twilio-routed cost across 6 verticals landed at $0.041/min, which is well under the $0.10/min margin floor we built into the pricing tiers ($149 / $499 / $1499).

The biggest single win came from caching system prompts across calls within a tenant — when the same tenant's salon receptionist takes 80 booking calls a day, the cache stays hot all day and average LLM cost dropped 67%. Try it on the 14-day no-card trial.

Optimization checklist

1. Amortize phone number cost across actual minutes — pick the right plan.
2. Always use Media Streams (cheaper than Twilio Conversation Relay on most workloads).
3. Use a cascaded stack on Twilio for cost-sensitive verticals.
4. Use end-to-end Realtime on Twilio for premium verticals.
5. Convert Twilio's mu-law 8kHz to PCM16 24kHz once at the bridge — never round-trip.
6. Disable recording for non-regulated calls — you save $0.0025/min.
7. Watch outbound country routing — international can blow up your bill.
8. Cache LLM system prompts hot across calls within a tenant.
9. Log every line item to a cost table so you catch drift early.
10. Re-quote Twilio every 6 months — prices and discounts move.

FAQ

Is Media Streams the cheapest way to get audio out of Twilio? Yes for AI agent use. Conversation Relay is more expensive because it bundles ConvAI features.

Can I run Twilio inbound + BYO Realtime in production? Yes — this is a standard pattern. You convert mu-law 8kHz to PCM16 24kHz at the bridge.

What about Twilio's own AI Assistants product? It is convenient but more expensive (bundled per-minute fee). DIY bridges win on cost.

Where do most teams blow their Twilio budget? International inbound numbers, recording storage, and forgetting to release unused phone numbers.

How does this compare to Vonage or Plivo? Plivo is ~30% cheaper on inbound but smaller global footprint. Vonage matches Twilio. CallSphere uses Twilio for breadth.

Sources

• Twilio Programmable Voice US Pricing — https://www.twilio.com/en-us/voice/pricing/us
• Twilio Pricing Overview — https://www.twilio.com/en-us/pricing
• Twilio Media Streams docs — https://www.twilio.com/docs/voice/media-streams
• Deepgram Pricing — https://deepgram.com/pricing