Voice AI Agents in 2026: Where the Latency and the Money Go

By Daniel Kim · Eight Pine Labs

Vapi publishes a median latency under 500 milliseconds. When Cekura pushed more than a thousand turns through it on a fixed stack, the measurement came back at 2.34 seconds.

Both numbers are real. They measure different things, which is the problem with every latency figure in this market. Vendors quote the part of the pipeline they control. Your caller hears all of it.

This guide works through the measured numbers behind voice AI agents in 2026. Where the milliseconds go, and why the model you spent a week choosing barely moves the invoice.

The short version

• Measured voice-to-voice latency on production calls runs 680 ms to 850 ms at the median, and stretches past 1,180 ms at p95.
• Pin every platform to the same components and the spread widens to 1.44 s through 3.16 s per turn. Orchestration is doing more work than the model.
• Humans register a pause at roughly 800 ms and treat the conversation as broken near 1,500 ms.
• The LLM costs about $0.003 per minute on a chained stack. Platform fees run $0.05 to $0.14+ per minute, so the model lands under 5% of the bill.
• Advertised platform pricing of $0.05/min becomes $0.23 to $0.33/min once every component is layered in.

What the measured latency numbers say

Two independent tests ran in 2026. They disagree, and the disagreement is instructive.

Tested Media sampled 500 production calls in March 2026, letting each platform run its own tuned stack, as compiled by Telnyx. Those results:

Cekura asked a different question. It pinned every platform to identical components (GPT-4.1, Deepgram Nova-3, ElevenLabs Flash) across more than a thousand turns, isolating what the orchestration layer itself contributes:

Retell sits near the top of both tables, which is a decent signal. The wider lesson is that a platform running its own optimized defaults behaves nothing like the same platform running components you chose. If you plan to bring your own models, the second table predicts your experience better than the first.

Set both against what the vendors publish:

The ElevenLabs figure is the one to watch out for. 75 milliseconds is a text-to-speech model latency, and it is accurate. It describes maybe a tenth of the round trip a caller experiences.

Where the milliseconds go

Latency in a voice agent is additive, and each hop is measurable. Telnyx published a component-level breakdown of a co-located stack against a typical stitched one:

Softcery models a target budget the same way and lands near 800 ms: voice activity detection around 50 ms, speech to text at 150 ms, LLM time-to-first-token at 400 ms, speech synthesis at 150 ms, network at 50 ms.

Notice which line dominates. LLM inference is 650 of the 1,210 milliseconds in the stitched stack, and 400 of Softcery's 800. Everything else combined is smaller than the model call. That makes time-to-first-token the number worth optimizing, and it explains why a fast small model beats a slow smart one for anything conversational.

The carrier leg is small and it is not zero. A June 2026 TECHSY test measured round-trip time at the carrier: Telnyx 71 ms p50 and 118 ms p95, Twilio 89 ms and 161 ms, Vonage 94 ms and 152 ms. Twenty milliseconds of carrier difference will not save a stack that spends 650 milliseconds waiting on a model.

How fast does a voice agent need to be?

Human conversation hands off turns at roughly 200 milliseconds. That is the target nobody hits.

The thresholds that matter operationally:

• Around 800 ms, listeners start to notice the pause.
• Around 1,500 ms, the conversation reads as broken and people start talking over the agent.
• ITU-T G.114, the telecom standard, caps one-way transmission at 150 ms before call quality degrades.

Line those up against the Cekura numbers and the picture gets uncomfortable. Every platform in that fixed-stack test sat above 1.4 seconds per turn. Four of the seven were past the 1,500 ms mark where callers give up on waiting. Production numbers look better because vendors tune hard, so the gap between a tuned deployment and a naive one is the difference between a usable product and one that gets hung up on.

The model is a rounding error

Here is where most teams spend their time backwards. The pattern will look familiar if you read how crypto became Wall Street’s R&D department, where the layer running the experiment rarely ends up collecting the fees.

A chained stack running Claude Haiku 4.5 at 1,500 input tokens and 250 output tokens per turn costs roughly $0.003 per minute in LLM tokens. Managed platform fees run $0.05 to $0.14+ per minute. The model you researched for a week is often under 5% of your bill.

Text model pricing for chained stacks, per million tokens, per Layer3 Labs:

Speech-to-speech models bill on audio tokens and cost considerably more:

The supporting components price out small. Deepgram streaming speech-to-text runs $0.0048 per minute. AssemblyAI charges $0.15 per hour. Deepgram's Aura-2 voice runs $0.030 per thousand characters, landing around $0.02 to $0.03 per spoken minute.

What voice agent platforms cost per minute

Advertised rates and real rates diverge by a factor of four to six once every component is billed:

A modular platform quoting $0.05 is quoting orchestration. Speech-to-text, the model, speech synthesis and telephony all arrive separately. Bundled pricing looks more expensive on the page and often is not once the invoice lands.

Where the cheap models change the math

DeepSeek's V4 pricing sits far enough below the frontier labs that it changes which workloads are worth automating at all.

Both carry a 1M token context window and up to 384K tokens of output, with thinking and non-thinking variants. V4 Flash undercuts GPT-5.5 at $5/$30 by roughly 99%, and sits below Claude Sonnet 4.6 at $3/$15. V4 Pro comes in under Claude Opus 4.7 at $5/$25.

Two details matter for voice specifically. Cache hits price at about 2% of a cache miss, and voice agents re-send a long system prompt on every single turn, which is close to the ideal caching pattern. From August 16, 2026 DeepSeek also bills dynamically, with off-peak rates at half of peak. Peak windows run 01:00 to 04:00 and 06:00 to 10:00 UTC, so batch and outbound campaigns can be scheduled around them.

The caveat is latency. Cheap tokens do nothing for time-to-first-token, and TTFT is the line that dominates the budget above. Price the model on cost, choose it on speed.

Which voice agent platform should you choose?

Softcery reviewed twelve platforms covering most of the market in 2026, and they sort cleanly by what you are optimizing for.

You want control over every component

Vapi, Pipecat, LiveKit, Ultravox and Cartesia Line expose the full stack. Pipecat reached v1.0.0 in April 2026 and LiveKit shipped v1.5.x through March and April, so the open-source options are past the point where you are debugging the framework instead of your product. Budget $0.23 to $0.33 per minute in reality.

You want to launch this week

Synthflow, Telnyx Voice AI Agents and NiCE Cognigy ship visual builders and bundled telephony. Launch time is measured in hours. You give up component swapping, which becomes the constraint about six months in.

You are doing outbound at volume

Bland handles voicemail detection, call transfer and DTMF as first-class features. Retell covers similar ground with better compliance posture.

You are in a regulated industry

SOC 2 Type II covers Vapi, Retell, Synthflow, Cartesia, ElevenLabs and Deepgram. HIPAA narrows to Vapi, Retell, Synthflow and ElevenLabs Conversational AI 2.0. PCI DSS covers Vapi, Telnyx, Ultravox and Retell. Compliance requirements shrink the shortlist faster than any technical criterion, so start there.

You need the best individual components

ElevenLabs leads on voice expressiveness and cloning. Deepgram leads on transcription accuracy with Nova-3. Cartesia's Sonic-3 runs under 100 ms of model latency for synthesis.

Cascaded or speech-to-speech?

Cascaded architectures run speech to text, then the model, then speech synthesis. Vapi, Retell, Synthflow, Deepgram and Cartesia work this way. You can see and swap every stage, and you pay for it in accumulated latency across four hops.

Speech-to-speech models skip the text round trip. LiveKit, Pipecat and ElevenLabs Conversational AI 2.0 support this path, and the audio token pricing above is what it costs. Debugging gets harder because there is no transcript at the boundary to inspect when the agent misunderstands someone.

Ultravox takes a third route with a GLM-4.6 backbone at 355 billion parameters and 160 experts per layer, scoring 87.05 on VoiceBench without reasoning enabled.

For most teams shipping in 2026, cascaded remains the safer default. You keep transcripts for evaluation and compliance, and the latency penalty is recoverable through a fast model and co-located components.

How to run this decision

1. Start with compliance. HIPAA or PCI requirements eliminate most of the field before you have benchmarked anything.
2. Set a latency budget before choosing components. Work backward from 800 ms, and treat LLM time-to-first-token as roughly half of it.
3. Benchmark on your own stack. Vendor p50 figures describe their tuned configuration. Cekura's numbers show what happens to yours.
4. Price the whole invoice. Multiply advertised modular pricing by four before comparing it to a bundled quote.
5. Choose the model last. It is under 5% of cost, and its only decisive property is time-to-first-token.

Frequently asked questions

What is a good latency for a voice AI agent?

Under 800 milliseconds voice-to-voice. Listeners begin noticing delay around 800 ms and treat the conversation as broken near 1,500 ms. Human turn-taking sits near 200 ms, which no current stack reaches.

Which LLM is best for voice agents?

Whichever fast model your platform supports. Claude Haiku 4.5, gpt-5.4-mini and Gemini 3.5 Flash all price between $0.75 and $1.50 per million input tokens and cost roughly $0.003 per minute in a chained stack. Time-to-first-token decides this, since the model contributes 400 to 650 ms of an 800 to 1,200 ms budget.

How much does a voice AI agent cost per minute?

Between $0.05 and $0.33 per minute depending on architecture. Bundled platforms run $0.05 to $0.14. Modular platforms advertise $0.05 and land at $0.23 to $0.33 once speech-to-text, the model, synthesis and telephony are billed separately.

Is speech-to-speech faster than a cascaded stack?

Yes, because it removes the text round trip between transcription and synthesis. It costs more, with gpt-realtime-2.1 at $32 per million audio input tokens against $1 per million text tokens for Haiku 4.5, and it gives up the transcripts most teams need for evaluation.

Why do vendor latency claims differ so much from measured results?

They measure different segments. ElevenLabs quoting 75 ms is describing one text-to-speech model. Bland's 400 ms has no published methodology. Independent fixed-stack testing put the same platforms between 1.44 and 3.16 seconds per turn, because that test measured the full turn on components the tester chose.

Should I build on open source or buy a managed platform?

Build on Pipecat or LiveKit when you need edge deployment or full visibility and have infrastructure engineers. Both hit production maturity in 2026. Buy Synthflow or Telnyx when speed to market matters more than component control. Buy Vapi or Retell when you want component flexibility without running infrastructure.

Eight Pine Labs advises companies building at the frontier of AI, fintech and blockchain. If you are scoping a voice agent deployment and want a second opinion on the stack, get in touch. More research is in the archive.

Sources

• Telnyx, Voice AI agents compared on latency. Tested Media production-call study (March 2026), Cekura fixed-stack benchmark, TECHSY carrier round-trip test (June 2026), component-level breakdown.
• Softcery, Choosing the right voice agent platform in 2026. Twelve-platform comparison, per-minute pricing, compliance matrix, latency budget.
• Layer3 Labs, Best LLM for voice agents. Model pricing for chained and speech-to-speech stacks, per-minute cost modelling.
• CostGoat, DeepSeek API pricing. V4 Flash and V4 Pro rates, caching behaviour, off-peak billing.
• ITU-T Recommendation G.114. One-way transmission time limits.