Griffin-Lite 10 ms audio packets vs Sume's 4 second minimum clip

The smallest unit Tavus describes for Griffin-Lite is an audio packet of 10 ms, generated one latent at a time. The smallest unit Sume Avatar 1.0 plans is a 4 second clip, inside a 4 to 60 second job. That gap, about 400 times, is the real difference between a live conversational model and a rendered talking video, and it decides which one you can use this week.

What Tavus says it is built for

The Tavus page describes a model that takes a reference image together with streaming audio and controls, and reports an average latency of 0.43 seconds. It also states that Griffin-Lite is not available to customers at this time and is open to select trusted testers as a research preview. There is nothing to integrate yet unless Tavus has invited you.

Tavus also says it is working on safe disclosure features before wider release, because a model people can mistake for a person needs them. Plan for that on any avatar you ship.

What the Sume timing looks like

A Sume talking-video request is a job. The job states are queued, processing, completed, failed and canceled, and the docs say queued is a normal accepted state. You submit with an Idempotency-Key, poll status or take a signed webhook, then read the result URL. Sync mode only blocks for a bounded wait and may still return a queued or processing job, so do not build a chat loop on it.

The 4 second floor is billed. At the standard tier of $0.184 per second, a 4 second clip is 4 x 0.184 = $0.736. At plus ($0.245) it is $0.98 and at max ($0.55) it is $2.20. The 60 second ceiling is 60 x 0.184 = $11.04 on standard.

Which problem each timing fits

Use a streaming model when a person talks to the avatar and expects it to react within a second. Use a rendered clip when you know the words in advance, want to review them, and want one file that any number of viewers can watch.

If you need both, a common split is a live front door for open questions and a bank of rendered clips for the answers you already know. That second part is something you can build now. The first part waits for access from the vendor.

Planning around the gap

Do not treat the two as competing for the same slot. A live model answers in the moment, so its value is the unplanned part of a conversation. A rendered clip is planned, so its value is control and reuse. Write down which parts of your flow are predictable, such as the greeting, the plan explanation and the next-step instruction, and which are not. Render the first group on Sume and leave the second to a person or a live product you have access to.

Also keep expectations on cost visibility. Sume publishes per-second rates, so you can compute a budget before any call. The Tavus page we read does not state a price, so a live comparison cannot be done on cost today.