Tavus PAL Maker, API, Enterprise vs the Sume avatar API: what matches

None of Tavus's three products maps one-to-one to Sume's avatar endpoints, because they solve different problems. Tavus's home page, read 2026-10-07, lists PAL Maker (no-code build and deploy of Personal AI Assistants), a Developer API called CVI, and bespoke Enterprise deployments. Sume's avatar API renders a scripted talking-video file from a reusable avatar. If you need something that talks back, the Tavus products are the relevant ones; if you need a finished clip, Sume is.

What the Tavus page says

The page names four models alongside the products: Sparrow-2 for turn-taking, Phoenix-4.5 for rendering, Raven-1 for perception, and Griffin, labelled as a preview for interaction. It quotes 134 ms for Phoenix-4.5 latency, under 300 ms context freshness for Raven-1, and says 48 percent thought Griffin was a real person on live video. It lists no prices on that page and points to a pricing section. These are vendor-stated figures, not Sume measurements, and this post does not test them.

What Sume covers

Sume creates a reusable avatar from a prompt, profile traits or a reference image, then renders a talking video from a script or ordered video_inputs of 4 to 60 seconds, with optional product and scene inputs, quality tiers, first-frame previews, inline captions and a beta face-swap endpoint. Each request creates a job that you poll or receive by webhook. There is no live session, no turn-taking and no viewer-side perception. Sume does not offer a live or real-time conversational avatar.

How to choose

• Viewer speaks and the avatar answers on the spot: the Tavus side of the table, or any live agent product. Sume does not do this.
• Same message to many people, reviewed before it ships: a rendered clip on Sume.
• A mix: a rendered clip answers the ten most common questions, and a live agent handles the long tail. The live-versus-job post compares those two shapes.

Questions to put to any vendor

Ask whether the thing you are buying is a session or a file, who reviews what it says, and what you can store. A file can be approved line by line and kept in your records. A live conversation produces speech you did not script, so the review shifts to guardrails and logs. Pricing differs too: sessions are measured in minutes of use, and Sume's avatar clips are measured in seconds of output at published rates.

Griffin is described by Tavus as a preview; check its current access terms on the vendor page before you plan around it. See the invite-only post for what to build while you wait.

A note on dates and claims

Vendor pages change. The figures above were on Tavus's home page on 2026-10-07 and are the vendor's own wording. Sume has not tested them, and a number such as 134 ms describes a model stage, not what a viewer experiences across a whole pipeline. If a decision depends on them, ask the vendor for the test conditions and run your own trial.