Providers and pricing

Ten vendors, ordered by how likely you are to reach for them rather than alphabetically. The catalogue lives in code and is returned live by list_avatar_providers — prefer asking it over trusting this table, which is a copy.

The table#

These prices are ordering hints, never the number in a client quote. They were researched in August 2026 and every one is a secondary source. Unpublished means unpublished — not free, and not to be guessed at.

The caveats, which are the useful part#

• LiveAvatar (HeyGen) deprecated avatar_persona in favour of voice_agent within four months of shipping it. Expect churn.
• Beyond Presence sells a dedicated speech-to-video SKU at roughly half its managed-agent price. Ask for that one.
• bitHuman claims CPU-only inference at 25fps and a self-hostable SDK. If that holds it is the cheapest sovereign path in the table — verify before betting on it.
• Simli was quoted at both $0.009 and $0.05 per minute by different sources and the discrepancy is unresolved. Confirm directly.
• Tavus is the most expensive here and is sold as a managed pipeline rather than raw audio-in, so the comparison is not like for like.

Available is not the same as listed#

{ "name": "list_avatar_providers", "arguments": { "available": true } }

With available: true you get only the vendors whose key is actually set in this environment. The full list will happily name one that 401s at call time, which is a failure that surfaces mid-conversation.

The same distinction on HTTP: GET /providers/avatar?available=true. Control plane API

Which to choose#

• Cheapest, if its claims hold: bitHuman, and it is the only one that might let you stop paying per minute altogether.
• Most established: LiveAvatar, with the churn caveat attached.
• Node worker: six of the ten ship a Node plugin. The worker runtime here is Python, so that column is informational rather than a constraint.

Validate the choice with GET /providers/avatar/<id> before a call is attempted rather than during one — an unknown provider is a 404 you can handle, and a 401 mid-call is one you cannot.