Tavus

Tavus should be evaluated as conversational infrastructure with a face, not as another text-to-video editor. Its Conversational Video Interface bundles perception, turn-taking, speech recognition, an LLM, text-to-speech, WebRTC and a rendered replica. That removes integration work, but it also means the video layer cannot be judged independently from session lifecycle, model behavior and concurrency.

> Distinctive strength: Modular live video-agent controls. > > Where it stops being an advantage: Live outcomes drive value and cost.

The decision in 60 seconds

Conversation and generated-video minutes are separate

The official pricing page lists Starter at $59 per month, with three custom Replica trainings each month, 100 conversational minutes, ten generated-video minutes and up to three concurrent streams. Growth is $397, with seven trainings, 1,250 conversational minutes, 100 generated-video minutes and up to ten concurrent streams on the reviewed plan. Extra custom Replicas are listed at $65 on Starter and $40 on Growth. Enterprise is quote-led.

The API overview says conversational billing begins when the Replica enters and waits in the room and ends when the conversation finishes or times out. That makes session hygiene a financial control. A user who abandons the page without a clean termination can consume minutes even when no useful conversation occurred.

At full included use, Growth's base subscription is about $0.318 per conversation minute before LLM, integration and support labour. At 500 useful minutes, it is $0.794. Utilisation changes the economics more than the sticker price.

Test an outcome, not a demo loop

Choose one narrow task: qualify a lead, explain a benefit, coach an employee or collect intake. Define what the agent may say, what it must refuse and when a human takes over. Run at least fifty sessions that include silence, interruptions, accent variation, camera denial, background noise, reconnection and unsupported questions.

Capture time from page open to first useful response, overlap errors, false visual interpretations, abandon rate, transfer success and billed session time. Review transcripts and recordings under an approved retention policy.

`cost per resolved session = subscription + overage + external models + engineering + review ÷ sessions meeting the outcome`

Do not use average session duration alone. A short failed conversation can look efficient while reducing conversion.

Modularity creates control and testing obligations

Tavus documents a full managed pipeline and Echo modes that can bypass perception, speech recognition or the LLM. That is valuable for teams with existing agent infrastructure. It also creates configuration combinations with different privacy, latency and failure behavior. Freeze the exact Persona, Replica, pipeline layers, LLM/TTS vendors and timeout settings for a benchmark.

The create-conversation API supports a test mode that returns an ended conversation without the Replica joining. Use it for integration checks that should not consume live capacity, but do not mistake it for a media-quality test.

Replica consent must remain revocable

Tavus's acceptable-use policy requires explicit informed consent for customer Replicas, maintenance of that consent, removal when the subject revokes it and prominent disclosure that end users are interacting with AI. The troubleshooting documentation also requires a spoken consent statement in training footage.

Store consent scope outside the vendor: brand, channels, languages, sensitive topics, start/end date and revocation owner. A trained Replica is an operational identity asset, not a reusable stock file.

Do not put Tavus in front of users when

• the product does not need visual interaction and a voice or text agent would be simpler;
• there is no reliable escalation, termination or subject-revocation workflow;
• the fifty-session pilot misses latency, completion or cost-per-resolution targets.

Better alternatives for specific buyers

Tavus verdict

Tavus is one of the stronger API-first candidates when the product itself needs face-to-face AI interaction. Its value comes from replacing a multimodal integration stack, not from inexpensive generated clips. Approve it only when a live pilot proves controlled turn-taking, reliable termination, safe escalation and sustainable cost per resolved session.

Primary action: Open Tavus's current conversational pricing ↗, freeze the pipeline configuration, and approve only after fifty consented sessions meet resolution, latency, retention and escalation gates.

Official sources

• Tavus plans, conversational minutes, generation and concurrency ↗
• Tavus CVI architecture ↗
• Tavus conversation lifecycle and billing start ↗
• Tavus create-conversation API and test mode ↗
• Tavus Echo Mode pipeline control ↗
• Tavus replica troubleshooting and spoken consent ↗
• Tavus acceptable-use, disclosure and revocation obligations ↗

• Tavus privacy, conversation-log and connected-service retention ↗

Sources checked 2026-08-31. Reconfirm overage and enterprise compliance in the signed order.