Skip to content

capability

verified July 2026

Serving that degrades before it fails.

A live call cannot buffer, retry quietly, or show a spinner. So the serving stack is built on one rule: overflow costs the caller latency, never availability — and when even that runs out, the refusal is fast and machine-readable.

01

The three tiers

Fig. — One endpoint, three strata

01Primary GPU fleet

takes all traffic first

the published figures: 107 ms first audio single-stream

02Warm spill

model loaded, idling until the fleet fills

overflow costs latency, not availability

03Independent fallback

a separate failure domain, always answering

the farthest tier — still audio, never silence

past absolute capacity: 503 at_capacity, retryable — never dead air on a live call

02

The refusal, in full

Past what all three tiers can serve, the API refuses fast instead of letting every stream degrade — a full lane never quietly slows the others. The refusal is a signal your orchestrator can act on:

what past-capacity actually returns
HTTP/1.1 503 Service Unavailable
retry-after: 1

{"error": "at_capacity", "retryable": true}

# a busy signal your orchestrator can act on —
# retry, spill to a standby line, reschedule the job.
# the one thing that never happens: dead air on a live call

03

How to hold us to it

The load benchmark publishes what one GPU does at sixteen callers — 430 ms median, 813 ms p95, zero failed requests — because a reliability claim without a load figure is a slogan. For your own posture, standby failover lines run the same discipline in the other direction: our capacity idling beside your current vendor at $20 a month.

04

Notes — an engineer's checklist

01What does a fleet outage look like from my side?

Latency, not errors: traffic lands on the warm spill tier, which is already loaded and answering health checks, and past it on the independent fallback. The request still returns audio, from a warmer or farther tier.

02Why refuse instead of queueing?

Because a queued live call is dead air — the one failure a caller cannot interpret. A fast retryable 503 gives your orchestrator something to act on inside the same second.

03Is the 503 counted against my lines?

No — a refused request is not a conversation. If refusals correlate with your own traffic spikes, that is the signal to raise the burst cap or add a committed line.

04Can I make my stack resilient to you?

Yes, and we sell the tool: standby failover lines hold tested capacity beside any primary vendor — including treating us as the primary — at $20 a month per line.

See also

Related sheets.

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

Hear it on your own script