Zazu — LLM inference for voice agents, billed per minute

LLM inference for voice agents

Other LLM endpoints see requests. Zazu sees the whole call: turns, interruptions, silences. Latency stays low all call. Your TTS starts sooner. And it costs less than per-token billing.

Why Zazu

Always respondsNo rate limits mid-callZazu reserves GPU capacity when a call starts. The call runs to the end with no rate limits.
Responds fasterLow latency all callZazu ensures low p95 latency, not p50, because even one slow turn breaks a conversation
Costs lessPay per minute, not per tokenStop paying extra for system prompts, barge-ins and pre-emptive generation.

Zazu API

Qwen3.5-9B served at 5× lower latency, measured on Vapi

5.3×
faster than GPT-4.1 mini: 1.4×
higher τ²-bench score

Read the Vapi test →

Turn 1“Hi, I'd like to book a general check-up for next Tuesday afternoon, please.”

Qwen3.5-9B on Zazu

1.00 s

GPT-4.1 mini

1.64 s

Turn 1 of the test call. Hatched: LLM request to first token.

Up to 40% cheaper, with per-minute billing

~95% of tokens are resends.The system prompt and the whole conversation, sent again every turn.
Per-token bills ride on cache hits.And you don't control the cache hit rate.
No pre-emptive generation needed.It cuts latency in LiveKit but adds up to 50% more tokens. Zazu is fast without it.
Drop-in setup.A LiveKit or Pipecat plugin* and a base-URL change.

* Own orchestrator? Onboard to Zazu in one Claude prompt (Yes, we will give it)

Get API access

40%
less than per-token billing

Sheet 08: one call, two billing methods. Illustrative.

Zazu dedicated

2.7× more calls per GPU, same model

Your model, your GPUs.For fine-tuned models or regulatory limits: the same Zazu stack, on GPUs dedicated to you.
Fewer GPUs for the same calls.Up to 2.7× more concurrent call streams per GPU than stock vLLM, at 27% lower p95 latency.
Same outputs.Zazu doesn't change what the model says, so your product behaves the same.
Simple pricing.A small management fee per call-minute + hardware costs.

Talk to us about dedicated

2.7×
call streams per GPU: −27%
p95 latency at the limit

Sheet 07: the same GPU, before and after. Illustrative.

Zazu: gearboxes for inference

Think of a model on a GPU as an engine. The inference stack around it is the gearbox.

Most endpoints have a gearbox with one gear. You either work the way it wants, or pay for an engine running badly.

Zazu starts from your goal. Admission, memory, scheduling and caching adapt to the work, so the metrics you care about hold and the engine stays in its best range.

Voice agents are our first ready-to-use gearbox.