Skip to content

capability

verified July 2026

Low latency, with the receipts.

Every vendor says low latency. This page is what we mean by it: the number, what it includes, and how to catch us if it drifts.

01

What the figure includes

Our 107 ms is server time to first audio on the production API — gateway, queue, and synthesis, measured under call-shaped load with cloned voices. It is not model-only time, not a staging box, and not a marketing rounding. The p95 is 108 ms: one millisecond of spread on a single stream. Under sixteen concurrent callers the pair is on the load benchmark, stated rather than hidden.

02

Audit it from your own logs

Every utterance returns its own receipt — the ttfa_ms field in the completion metadata. Log it and the latency claim becomes your dashboard, not our marketing.

the receipt every utterance carries
// the receipt every utterance carries, on every surface
{"ttfa_ms": 107, "audio_ms": 2810}

# log ttfa_ms per utterance, alert on the tail:
#   single stream, production API — 107 ms p50 · 108 ms p95
#   what your alert should ask: has MY p95 moved?
# if your histogram disagrees with this page, bring it
# to a call — the harness reruns in front of you

03

Why the turn budget cares

Humans hand a conversation back and forth in about 200 ms and interrupt an agent much past a second. Synthesis is the one stage of that budget you can buy down outright — spend 107 ms here and the rest of the second belongs to your LLM. The full stage-by-stage budget is drawn on the customer-support page.

04

Notes — an engineer's checklist

01Is 107 ms the number I will see end to end?

No — it is server time, so your network round trip sits on top. That is stated on every figure because it is the only honest way to publish latency: we measure what we control, you add what you control.

02What does load do to the figure?

Sixteen concurrent callers on one GPU move the pair to 430 ms p50 / 813 ms p95 with zero failed requests — published on the load benchmark with the setup stated. Load costs latency here, not errors.

03How is the figure kept honest over time?

The harness runs against the production endpoint and reruns live on any demo call. Every utterance you synthesize also returns ttfa_ms, so your own logs audit the claim continuously.

See also

Related sheets.

Your script, this API, thirty minutes — and the recording leaves with you.

Hear it on your own script