Real-time AI avatar APIs compared: LiveAvatar, Tavus, Simli, Anam, and D-ID

Two kinds of avatar API

Batch APIs render a finished video from a script or an audio file. You wait seconds to minutes and get an MP4. HeyGen video API and Hedra Character-3 work this way. HeyGen switched that video API to pay-as-you-go credits in late October 2026; this page does not quote those rates.

Real-time APIs hold a live session. The user speaks, the agent answers, and the face moves while the answer plays. Latency decides whether the conversation feels natural, so vendors publish latency figures. The figures below are each vendor's own claims. Nobody has run them side by side on this page.

LiveAvatar (HeyGen)

LiveAvatar is HeyGen's real-time product, formerly Interactive Avatar. In Full mode it runs the voice, the language model, and the face. In Lite mode (Avatar Only on the LiveAvatar site) you bring your own LLM and voice and pay half the credits. HeyGen says the median time to first frame is under 300 ms. Plans start with 10 free credits; paid Essential and Business tiers add 1,100 and 6,000 credits. Full mode uses 2 credits per minute and Lite uses 1. See https://www.liveavatar.com and https://docs.liveavatar.com/docs/faq/credits for current rates. Sources: https://www.liveavatar.com , https://docs.liveavatar.com/docs/faq/credits , https://help.heygen.com/en/articles/12758516

Tavus Conversational Video Interface

Tavus CVI runs the full pipeline: perception (Raven), turn-taking (Sparrow), speech-to-text, an LLM you can replace, text-to-speech, and the Phoenix face model. Sessions run over WebRTC through Daily. Tavus says Phoenix-4.5 turns audio into video in 134 ms. It does not publish a single full-pipeline millisecond figure on the CVI or pricing pages reviewed for this article. The public pricing page lists two developer ladders at once (November 2026): Basic, Starter, and Growth at 25, 100, and 1,250 conversational minutes with overage on paid tiers, and Free, Starter, Builder, Growth, and Business at 20, 60, 175, 1,300, and 4,000 minutes. Use https://www.tavus.io/pricing rather than a copied table. Tavus lists SOC 2, HIPAA, GDPR, and the EU AI Act. Sources: https://www.tavus.io/cvi , https://docs.tavus.io , https://www.tavus.io/pricing

Simli

Simli renders the face only. You run the voice agent (often with LiveKit or Pipecat), send Simli the agent audio, and get video back. Simli says audio-to-video takes about 150 ms. Pricing is per minute on Hobby (1,000 minutes), Pro, Scale (27,500 minutes), and enterprise tiers, with 50 free minutes a month. See https://www.simli.com/pricing for current rates. Sources: https://www.simli.com , https://www.simli.com/pricing , https://docs.simli.com

Anam

Anam runs the face and can run the conversation, and it accepts your own LLM. It lists integrations with LiveKit, Agora, and ElevenLabs, over 70 languages, and SOC 2 Type II. Anam published its own benchmark in October 2026 with a 1.24 s median reply time for the full pipeline. Plans run from 30 free minutes through named Starter, Explorer (250 minutes), and Professional (8,000 minutes) tiers. See https://anam.ai/pricing for current rates. Sources: https://anam.ai/pricing , https://anam.ai/api , https://anam.ai/tools/latency-benchmark

D-ID

D-ID V4 Expressive Visual Agents answer from documents you upload (retrieval over your files) and stream over WebRTC. D-ID's tech specs say conversational latency is under 500 ms. Studio and API plan prices move, so use https://www.d-id.com/pricing rather than a copied table. Sources: https://www.d-id.com/v4-expressive-visual-avatars-tech-specs/ , https://www.d-id.com/news/v4-expressive-visual-agents-real-time-llm-connected-interaction/ , https://docs.d-id.com

How to choose

Pick from what you already run.

Whichever API renders the face, your app still owns the session records, transcripts, recordings, and anything the avatar should remember next time. None of these vendors is a database for that data.

• If you have no voice agent yet and want one vendor, start with LiveAvatar Full mode or Tavus CVI.
• If you already run a LiveKit or Pipecat agent, a face-only renderer such as Simli, or LiveAvatar Lite, keeps your LLM and voice choices.
• If the avatar answers from a document set and you want that hosted, D-ID agents include retrieval.
• If you need many languages or SOC 2 Type II paperwork, check Anam and Tavus first.
• If you only need pre-rendered clips, a batch API (HeyGen video API, Hedra Character-3) is usually cheaper per finished second than holding a live session. Check current batch rates; HeyGen API credits changed in late October 2026.

Last reviewed

Facts on this page were checked against vendor documentation on 2026-11-02. Product surfaces move; treat version names and APIs as directional and follow the official links in Related.

Real-time AI avatar APIs (last reviewed: Nov 2026)

Where Pixeltable fits

Use Pixeltable when transcripts, session recordings, generated clips, and per-user memory should live in tables you query, with embeddings kept in sync as rows arrive. Pixeltable does not render faces. Keep the avatar API you chose and post each finished turn to a table. The table embeds it, and a query route returns what this user said in earlier sessions so your agent can put it in the prompt.

import pixeltable as pxt
import pixeltable.functions as pxtf
from pixeltable.functions import openai
from pixeltable.serving import FastAPIRouter

TableModel = pxt.model_base()
embed = openai.embeddings.using(model='text-embedding-3-small')


class Turns(TableModel, name='turns'):
    id = pxt.Column(value=pxtf.uuid.uuid7(), primary_key=True)
    user_id: pxt.String
    session_id: pxt.String
    speaker: pxt.String  # 'user' or 'avatar'
    text: pxt.String

    __indexes__ = [pxt.EmbeddingIndex(text, embedding=embed), pxt.BtreeIndex(user_id)]


@pxt.query
def recall(user_id: str, query: str):
    sim = Turns.text.similarity(string=query)
    return (
        Turns.where(Turns.user_id == user_id)
        .order_by(sim, asc=False)
        .limit(5)
        .select(Turns.session_id, Turns.speaker, Turns.text)
    )


api = FastAPIRouter(name='avatar')
api.add_insert_route(Turns, path='/turns', inputs=[Turns.user_id, Turns.session_id, Turns.speaker, Turns.text], outputs=[Turns.id])
api.add_query_route(path='/memory', query=recall)

Questions