Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Anam / Cara

Cara 3

Previous stable face-generation model with 720×480 landscape output for compatible existing avatars.

availableClosed weightsVideo generationAudio-driven video
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
6 of 18 specification fields collected
Output resolution720×480 landscapeSee specification notes below
Clip durationNot collectedAwaiting source review
Input billingSee ratesAnam
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

No exact model-version contract established; missing information does not mean unsupported.

Maximum output
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Context caching
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Reasoning
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Tool calling
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Structured output
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Batch processing
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Fine-tuning
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Parameters
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Architecture
Audio-to-motion diffusion transformer with flow matching, followed by a separate motion-to-face rendererSource for Architecture (opens in a new tab)
Knowledge cutoff
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Languages
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Release date
Not collected

No exact model-version contract established; missing information does not mean unsupported.

Input modalities
Image, AudioSource for Input modalities (opens in a new tab)

Reference identity and supplied speech drive the visual layer. Cara 4 text controls direct performance; the surrounding LLM and TTS are separate.

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

Synchronized supplied audio is not proof of native speech generation.

Output resolutions
720×480 landscapeSource for Output resolutions (opens in a new tab)
Provider-reported latency
About 70 ms time to first generated frame on an H200Source for Provider-reported latency (opens in a new tab)

Research generator timing, not an end-to-end speech-to-response guarantee.

Access & pricing

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

API model ID
cara-3
Access
POST https://api.anam.ai/v1/auth/session-token with personaConfig.avatarModel=cara-3 and a compatible avatarId. Tokens are bound to the supplied configuration and valid for one hour.
Regions
Production inference/media on CoreWeave in Spain/EU and US East. Enterprise can select eu/us engine regions with strict or preferred routing. External providers and stored assets are separate.
avatarCreation
Existing compatible avatars remain usable. New Cara 3 creation is no longer supported.Source for avatarCreation (opens in a new tab)
directorNotes
Ignored on older models; configuring Director Notes does not add Cara 4 behavior to Cara 3.Source for directorNotes (opens in a new tab)
Billing basis
Billed per second of active session duration, including silence and idling. Monthly included minutes do not roll over. External LLM/TTS charges can be separate for bring-your-own pipelines.Source for Billing basis (opens in a new tab)
sessionLimits
Free 3 minutes; Starter 5; Explorer 10; Growth and Professional 2 hours; Enterprise unlimited under the published plan table. Account agreements and overrides govern.Source for sessionLimits (opens in a new tab)
Concurrent requests
Free/Starter 1; Explorer 3; Growth 5; Professional 10; Enterprise custom, advertised from 100 to unlimited. Shared across the organization.Source for Concurrent requests (opens in a new tab)
subscriptions
Starter $12/month with 50 included minutes; Explorer $49 with 250; Growth $299 with 2,000; Professional $999 with 8,000. Free has 30 and no overage.Source for subscriptions (opens in a new tab)
enterpriseRate
Advertised from $0.04 per minute at scale, subject to negotiated Enterprise terms; not a universal self-serve rate.Source for enterpriseRate (opens in a new tab)
regionalRouting
Enterprise sessionOptions.region: eu or us. Strict routing fails rather than crossing region; preferred routing can fail over. Engine selection does not govern all stored data or external providers.Source for regionalRouting (opens in a new tab)
retentionControls
Enterprise zeroDataRetention applies to session content. Operational infrastructure logs retained 30 days; avatar source images and uploaded knowledge are excluded. Not every LLM/voice supports ZDR.Source for retentionControls (opens in a new tab)

Prices retain their published currencies and billing units below.

Published billing units · original currency · observed 14 Oct 2026
CategoryUsageRateUnit
Session overageStarter session usage1$0.16minute of active session
Session overageExplorer session usage1$0.14minute of active session
Session overageGrowth session usage1$0.12minute of active session
Session overageProfessional session usage1$0.11minute of active session

Rate conditions

  1. Paid plan and included allowance apply. Billed per second; active silence and idle avatar time count. Minutes are aggregate session usage, not additive charges for face/STT/LLM/TTS. External providers can bill separately. Experimental model access and compatible avatar assets must be checked; no token-price conversion.
Pricing source (opens in a new tab)

Data & enterprise terms

Default recordings retained up to 30 days; reports/transcripts/analytics remain until deleted. Enterprise ZDR keeps session content in volatile memory but excludes source avatar images, knowledge files and operational/account metadata. Applicable terms prohibit customer-content training, subject to any separate written agreement described in governance.

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.

Anam

cara-3
Not verified

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

How it works
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

Minimum reusable input
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

Lifetime & refresh
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

How to enable it
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

Usage fields
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

Billing rules
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

Limits & data boundaries
Not collected

No provider prompt-cache API or model-specific cached tariff has been established. Saved avatars and internal inference state are separate.

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

Caching reviewed 14 Oct 2026

What the measurements say

Generator time to first frame

Anam · provider-reported
70ms

Provider research reports approximately 70 ms for its two-stage model on an H200. Excludes user transcription, LLM, TTS and network/session startup.

Evaluation source (opens in a new tab)

Lip-sync rating

Anam · provider-reported
7.5provider-reported rating

Values in the Cara 3 provider article. Linked Mabyduck methodology describes an Anam-commissioned October 2025 study, 178 participants and a 5-point Likert scale, not an exact checkpoint-specific validation of these displayed scores. Scale conversion and later study details remain unresolved.

Evaluation source (opens in a new tab)

Naturalness rating

Anam · provider-reported
6.3provider-reported rating

Values in the Cara 3 provider article. Linked Mabyduck methodology describes an Anam-commissioned October 2025 study, 178 participants and a 5-point Likert scale, not an exact checkpoint-specific validation of these displayed scores. Scale conversion and later study details remain unresolved.

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

Anam’s governance page explicitly identifies CARA as proprietary avatar models it develops and owns. Hosted access and open SDKs do not grant access to those model weights.

Classification evidence (opens in a new tab)

Lifecycle dates

Availability reported in changelog29 Dec 2025Source (opens in a new tab)
Research and public release article18 Feb 2026Source (opens in a new tab)

Sources reviewed