Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Tavus / Phoenix

Phoenix-4.5

Full-frame upper-body renderer with immediate zero-shot previews followed by background identity refinement.

Video generation
availableClosed weightsVideo generationAudio-driven video
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
7 of 17 specification fields collected
Output resolutionNot collectedAwaiting source review
Clip durationNot collectedAwaiting source review
Input billingSee ratesTavus CVI
Output billingSee ratesPublished units and conditions below

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Maximum output
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Context caching
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Reasoning
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Tool calling
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Structured output
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Streaming
SupportedSource for Streaming (opens in a new tab)
Batch processing
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Fine-tuning
SupportedSource for Fine-tuning (opens in a new tab)

Per-identity face refinement; not a public general-purpose weight fine-tuning endpoint.

Parameters
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Architecture
Streaming diffusion-transformer animator and GAN-based full-frame rendererSource for Architecture (opens in a new tab)
Knowledge cutoff
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Languages
Not collected

Exact model-version information has not been established; unknown does not mean unsupported.

Release date
2026-09-10Source for Release date (opens in a new tab)
Input modalities
Image, Video, AudioSource for Input modalities (opens in a new tab)

Image/video establishes identity; supplied audio drives animation.

Output modalities
VideoSource for Output modalities (opens in a new tab)

Speech synthesis is a separate pipeline stage, not proof that Phoenix is a native TTS model.

Provider-reported latency
134 ms from audio input to rendered videoSource for Provider-reported latency (opens in a new tab)

Provider measurement, not end-to-end conversation latency or an SLA.

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Tavus CVI

Provider source (opens in a new tab)
API model ID
phoenix-4.5
Access
POST https://tavusapi.com/v2/faces with model_name=phoenix-4.5; use the resulting face in the conversational pipeline.
Regions
Privacy policy describes US processing and possible transfers to service providers in other countries. Exact model-serving location is not established.
trainingVideo
30 seconds speaking followed by 30 seconds listening in one continuous recording; minimum 25 fps and 1080p, maximum 750 MB; MP4 or WebM with H.264 video and AAC audio.Source for trainingVideo (opens in a new tab)

Training source requirements, not generated output limits.

trainingAccess
Supply one public/presigned training URL accessible for at least 24 hours after submission. Image training requires a stock voice_name or default_voice_id, not both. Video training derives a voice from the recording.Source for trainingAccess (opens in a new tab)
trainingPaths
Image or video; zero-shot preview in about a minute for images, or minutes for video, followed by background refinement.Source for trainingPaths (opens in a new tab)
Reference image limits
JPG/PNG, at least 512×512; one adult human-shaped subject. Chest-up or waist-up framing; full body, animals and non-human mascots are not supported.Source for Reference image limits (opens in a new tab)
refinementStatus
status=completed means preview usable. finetune_status=started still has a watermark; completed removes it; errored leaves a watermarked preview. status=error means unusable.Source for refinementStatus (opens in a new tab)
Billing and deployment scope
CVI bundle includes Phoenix-4.5, Raven-1, Sparrow-2, LLM, TTS and WebRTC. These are shared bundle charges, not standalone or additive component prices.Source for Billing and deployment scope (opens in a new tab)
sessionLimits
Table: Free 5 minutes, Starter 10, Builder 15, Growth/Business 60; Enterprise custom.Source for sessionLimits (opens in a new tab)

The Growth hero says no duration limit, conflicting with the table. Confirm contractual limits.

Concurrent requests
Free/Starter 1; Builder 3; Growth 10; Business 15; Enterprise custom.Source for Concurrent requests (opens in a new tab)
Subscription eligibility
Starter $22/month with 60 minutes and no overage; Builder $59 with 175; Growth $397 with 1,300; Business $975 with 4,000; Enterprise custom.Source for Subscription eligibility (opens in a new tab)

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
CVI bundle overageBuilder full conversation pipeline1$0.35minute of CVI session
CVI bundle overageGrowth full conversation pipeline1$0.31minute of CVI session
CVI bundle overageBusiness full conversation pipeline1$0.26minute of CVI session

Rate conditions

  1. Subscription and included allowance apply. Shared price for the complete pipeline; do not add component rates together. Each conversation has a 30-second minimum, rounded to the nearest 6 seconds. A created session can be billed even if no participant joins. Audio-only has the same current price. No text-token conversion.
Pricing source (opens in a new tab)

Data & enterprise terms

Default privacy terms permit service improvement and AI research/training. Biometric retention is limited by the stated purpose or one year after last interaction, whichever comes first. Enterprise zero-retention arrangements are separate; no universal zero-retention promise is inferred.

Read the service policy (opens in a new tab)

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

Tavus CVI

phoenix-4.5
Not verified

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

How it works
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

Minimum reusable input
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

Lifetime & refresh
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

How to enable it
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

Usage fields
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

Billing rules
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

Limits & data boundaries
Not collected

Internal streaming state, saved faces and conversation memories do not establish billable prompt caching, cache lifetime or discounts.

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 14 Oct 2026

What the measurements say

AVSR

Tavus · provider-reported
0.304score

Same paired internal test set; both models fine-tuned. Provider evaluation, not an independent cross-provider ranking. AVSR/CSIM higher is better; FID/FVD lower is better; ΔP describes motion vividness.

Evaluation source (opens in a new tab)

CSIM

Tavus · provider-reported
0.912score

Same paired internal test set; both models fine-tuned. Provider evaluation, not an independent cross-provider ranking. AVSR/CSIM higher is better; FID/FVD lower is better; ΔP describes motion vividness.

Evaluation source (opens in a new tab)

Motion vividness ΔP

Tavus · provider-reported
0.605score

Same paired internal test set; both models fine-tuned. Provider evaluation, not an independent cross-provider ranking. AVSR/CSIM higher is better; FID/FVD lower is better; ΔP describes motion vividness.

Evaluation source (opens in a new tab)

FID

Tavus · provider-reported
25.42score

Same paired internal test set; both models fine-tuned. Provider evaluation, not an independent cross-provider ranking. AVSR/CSIM higher is better; FID/FVD lower is better; ΔP describes motion vividness.

Evaluation source (opens in a new tab)

FVD

Tavus · provider-reported
148.48score

Same paired internal test set; both models fine-tuned. Provider evaluation, not an independent cross-provider ranking. AVSR/CSIM higher is better; FID/FVD lower is better; ΔP describes motion vividness.

Evaluation source (opens in a new tab)

Human preference Elo · zero-shot

Tavus · provider-reported
1002Elo

Provider head-to-head naturalness comparisons; evaluator count and calibration not disclosed in the reviewed article.

Evaluation source (opens in a new tab)

Human preference Elo · fine-tuned

Tavus · provider-reported
1069Elo

Provider head-to-head naturalness comparisons; evaluator count and calibration not disclosed in the reviewed article.

Evaluation source (opens in a new tab)

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Named model developed by Tavus in its proprietary hosted/research stack. Reviewed primary documentation provides no downloadable weights for this model; service terms restrict source derivation and independent avatar use. Open SDKs and third-party models in the surrounding pipeline are separate.

Classification evidence (opens in a new tab)

Lifecycle dates

Named release announced10 Sept 2026Source (opens in a new tab)

Sources reviewed