Back to the blog

November 18, 2026 · 12 min · api · education · engineering

.md

Max avatar call length by plan: what the hard stop does, end well

Longest call per avatar API plan, Nov 2026: Realtime Avatar 30 min on every plan from $9, Anam 5-120 min, Tavus 10-60, LiveAvatar 20-60. How to continue.

Realtime Avatar caps any single call at 1,800 seconds (30 minutes) through max_session_seconds, a hard stop set in each call request rather than a plan tier, so a call on the $9 Starter plan can run as long as one on the $119 Studio plan. Other entry plans that allow 25 to 30 minute calls, as of November 2026: LemonSlice Starter ($8 a month, 30-minute calls), Beyond Presence Starter (49 euros a month, session duration listed as unlimited) and Talking Avatar's pay-as-you-go Voice Agents ($0.06 a minute, calls up to 60 minutes). Anam ends calls at 5 minutes on Starter and 10 on Explorer, Tavus at 10 on Starter and 15 on Builder, and HeyGen LiveAvatar at 20 on Essential. Past any cap, start a new call and replay the recent conversation as context: Realtime Avatar accepts up to 32 prior messages per call.

Which plan to pick, by the call length you need

  • 25 to 30 minutes per call on an entry plan: Realtime Avatar Starter ($9 a month, 120 minutes, then $0.095 a minute) or Developer ($24, 600 minutes, then $0.085); LemonSlice Starter ($8 billed monthly, 41 minutes, then $0.22 a minute, plus $0.09 a minute when LemonSlice hosts the speech recognition, LLM and voice); Talking Avatar Voice Agents (no monthly fee, $0.06 a minute, a no-code agent built in its dashboard).
  • Up to 60 minutes per call: Talking Avatar Voice Agents on pay as you go, Tavus Growth ($397 a month, 1,300 minutes) and HeyGen LiveAvatar Business ($475 a month, 3,000 minutes in FULL mode).
  • Up to 2 hours per call: Anam Growth ($299 a month, 2,000 minutes) and LemonSlice Ultra ($960 a month billed monthly).
  • Session duration listed as unlimited: Beyond Presence's paid conversational video agent plans, from 49 euros a month for 140 minutes.
  • Longer than any per-call cap: close the call on a clean line and start a new one that replays a running summary plus the latest turns. Realtime Avatar accepts up to 32 prior messages per call.

Competitor figures below come from each vendor's own pricing page, read on 16 November 2026 and linked at the end; Realtime Avatar figures come from its pricing page. After the plan list: a 25-minute tutoring cost table, how the hard stop works, and code that carries one conversation across calls.

Max call length by vendor and plan, as of November 2026

Included minutes and call length are separate limits: a plan can include hundreds of minutes and still end every call at 10. Each line gives monthly price, included minutes, the rate after them, concurrency and the longest single call.

  • Realtime Avatar Sandbox: free, 480 credits once per account, no card (8 realtime minutes, or a first avatar plus 2 minutes of calls), 2 concurrent calls, 1 avatar.
  • Realtime Avatar Starter: $9 a month, 120 minutes included, then $0.095 a minute, 3 concurrent calls, 30-minute hard stop.
  • Realtime Avatar Developer: $24 a month, 600 minutes, then $0.085 a minute, 10 concurrent calls, 30-minute hard stop.
  • Realtime Avatar Studio: $119 a month, 3,000 minutes, then $0.08 a minute, 20 concurrent calls, 30-minute hard stop. Scale is quoted by sales.
  • Anam Free: $0, 30 minutes a month with no extra minutes, 1 simultaneous session, 3-minute conversations (anam.ai/pricing).
  • Anam Starter: $12 a month, 50 minutes, then $0.16 a minute, 1 simultaneous session, 5-minute conversations.
  • Anam Explorer: $49 a month, 250 minutes, then $0.14 a minute, 3 simultaneous sessions, 10-minute conversations.
  • Anam Growth and Professional: $299 a month for 2,000 minutes then $0.12 a minute, or $999 for 8,000 then $0.11; 5 and 10 simultaneous sessions; 2-hour conversations.
  • Tavus Free: 20 minutes of Conversational Video a month, 5-minute conversations (tavus.io/pricing).
  • Tavus Starter: $22 a month, 60 minutes, no pay-as-you-go overage, 1 concurrent session, 10-minute conversations.
  • Tavus Builder: $59 a month, 175 minutes, then $0.35 a minute, 3 concurrent sessions, 15-minute conversations.
  • Tavus Growth and Business: $397 a month for 1,300 minutes then $0.31 a minute, or $975 for 4,000 then $0.26; 10 and 15 concurrent sessions; 60-minute conversations.
  • HeyGen LiveAvatar Free: 10 credits a month, 1 concurrent session, 2-minute sessions, watermark (liveavatar.com).
  • HeyGen LiveAvatar Essential: $99 a month, 1,100 credits, which is 550 minutes in FULL mode (the managed agent, 2 credits a minute), then $0.095 a credit or $0.19 a FULL minute, unlimited concurrency, 20-minute sessions.
  • HeyGen LiveAvatar Business: $475 a month, 6,000 credits or 3,000 FULL minutes, then $0.09 a credit or $0.18 a FULL minute, unlimited concurrency, 60-minute sessions.
  • Talking Avatar Voice Agents: no monthly fee, $0.06 a minute all-in (avatar, AI model, speech recognition and voice), billed per second, with each agent's maximum call length set from 1 to 60 minutes. It is a no-code agent built in their dashboard and shared by link, widget or iframe; watermark removal is $19 a month (talking-avatar.dev/pricing).
  • Talking Avatar Realistic Avatars API: $0.02 a minute for the face alone, with your own AI model and voice, and sessions up to 3 hours.
  • LemonSlice Starter: $8 a month billed monthly ($7 billed yearly), 41 free minutes a month, then $0.22 a minute, concurrency limit 3, 30-minute conversation limit. Those rates cover the avatar with your own LLM or voice model; hosted avatars, where LemonSlice runs speech recognition, the LLM and the voice, cost $0.09 a minute more. Its Ultra plan ($960 a month billed monthly) allows 2-hour calls (lemonslice.com/pricing).
  • Beyond Presence Starter for conversational video agents: 49 euros a month, 140 included minutes, then 0.35 euros a minute, 10 concurrent sessions, maximum session duration listed as unlimited, against 3 minutes on Free. Its face-only Speech-to-Video API has its own allowance, 280 minutes on Starter (beyondpresence.ai/pricing).

Two notes. Tavus's page still shows an older plan block (a $59 Starter with 100 minutes, a $397 Growth with 1,250); the figures above are from its current plan cards and comparison table. LiveAvatar's LITE mode, with your own speech recognition, LLM and voice, costs 1 credit a minute under the same session limits.

What a month of 25-minute tutoring sessions costs

A tutoring app running 100 sessions of 25 minutes a month uses 2,500 minutes. Each vendor is priced on the plan that costs least at this volume among its self-serve plans that allow a 25-minute call: monthly price plus minutes beyond the allowance times the overage rate. Dollar figures come lowest first, then Beyond Presence, which prices in euros:

  • Realtime Avatar Studio: $119. The 3,000 included minutes cover all 2,500.
  • Talking Avatar Voice Agents: 2,500 × $0.06 = $150, or $169 with the $19 watermark removal.
  • Realtime Avatar Developer: $24 + 1,900 × $0.085 = $185.50.
  • Realtime Avatar Starter: $9 + 2,380 × $0.095 = $235.10.
  • Anam Growth: $299 + 500 × $0.12 = $359. Explorer is ruled out by its 10-minute cap.
  • HeyGen LiveAvatar Business in FULL mode: $475, with 3,000 FULL minutes included. Essential is ruled out by its 20-minute cap.
  • Tavus Growth: $397 + 1,200 × $0.31 = $769. Builder is ruled out by its 15-minute cap.
  • Beyond Presence Scale, in euros and so not ranked against the dollar figures: 349 + 500 × 0.175 = 436.50 euros. Its Starter (875 euros) and Growth (500 euros) cost more at this volume.

At this volume Studio is the lowest figure in the list. By the same formula, Developer costs less than Studio below about 1,718 minutes a month, and Starter less than Developer below about 278. If sessions overlap, check concurrency: Starter runs 3 calls at once, Developer 10, Studio 20.

Per session, once the included minutes are used up, one 25-minute call costs about $2.38 on Realtime Avatar Starter, $2.13 on Developer and $2.00 on Studio, $1.50 on Talking Avatar Voice Agents, $3.00 on Anam Growth, $4.50 in HeyGen LiveAvatar Business FULL mode and $7.75 on Tavus Growth: 25 times each plan's per-minute overage rate.

Where others fit better: if a no-code dashboard agent suits you, Talking Avatar's Voice Agents give hour-long calls at $0.06 a minute. Face-only renderers cost less per minute again: Talking Avatar's Realistic Avatars API is $0.02 a minute with sessions up to 3 hours, and Simli's Hobby plan is $10 a month for 1,000 minutes, with no per-call maximum stated on its pricing page. With either, you bring and pay for your own speech recognition, LLM and voice. On Realtime Avatar, credits also cover generation (images, idle clips, avatars), so building avatars draws on the same allowance. LemonSlice is left out of the cost list because the page read for this article shows included minutes for its Starter plan only.

How the hard stop works

On Realtime Avatar the limit is a field in the session policy your server returns when it mints a call: maxSeconds in the SDK, max_session_seconds on the wire, up to 1800. The docs describe it as a hard stop enforced on the platform's side, so it holds even if your client stops reporting, and they tell you to compute it from the balance you just admitted. Every field in that policy is authoritative: whatever the browser sent for it is discarded.

None of this depends on the plan, which sets included minutes, overage, concurrency and avatar count; the 1,800-second ceiling is the same on all of them. What can shorten a call is the maxSeconds you send and your workspace balance (GET /credits/balance returns what the next call must fit inside). In the browser, call.secondsRemaining counts down to the hard stop, and a call that reaches it ends with the reason session_cap, as listed in the Calls docs. A Next.js route that mints a capped, continuing call:

// app/api/realtime-avatar/[...path]/route.ts
import { createRealtimeAvatarRoute } from "realtime-avatar/nextjs";
import { buildContext } from "@/lib/lesson-context"; // shown below
import { loadLesson, admittedSeconds } from "@/lib/lessons"; // your own code

const HARD_STOP_SECONDS = 1800; // the most one call may last, on every plan

export const { GET, POST } = createRealtimeAvatarRoute({
  apiKey: process.env.REALTIME_AVATAR_API_KEY!,
  session: async ({ request, avatarId }) => {
    const lesson = await loadLesson(request, avatarId); // your auth and lookup
    const seconds = await admittedSeconds(lesson.learnerId); // what this learner may spend now
    if (seconds < 60) return new Response("Out of minutes", { status: 402 });
    return {
      instructions: lesson.tutorInstructions,
      context: buildContext(lesson.summary, lesson.turns),
      maxSeconds: Math.min(HARD_STOP_SECONDS, seconds),
      transcript: { url: "https://your.app/api/rta/transcript", secret: process.env.RTA_TRANSCRIPT_SECRET! },
      metadata: { lessonId: lesson.id },
    };
  },
});

A backend in another language sends the same policy as one POST, with the wire names:

POST /realtime/livekit/session   (wire names: snake_case and strict)
{
  "avatar_id": "YOUR_AVATAR_ID",
  "instructions": "You are a patient Spanish tutor for adult beginners.",
  "initial_context": [
    { "role": "system", "content": "Summary so far: covered ser vs estar; struggles with irregular past tense." },
    { "role": "assistant", "content": "Let's try exercise 4: describe your last weekend." },
    { "role": "user", "content": "El sabado yo fui al mercado." }
  ],
  "max_session_seconds": 1800,
  "transcript_webhook": { "url": "https://your.app/api/rta/transcript", "secret": "a-16-to-200-char-secret" },
  "client_metadata": { "lesson_id": "L-42" }
}

Continuing a conversation past the cap

A new session does not automatically inherit memory from the last one, so on Realtime Avatar continuity is something your server supplies. That is a real difference from Tavus, whose pricing page lists Memories, persistent memory across conversations, as a feature; here you choose what carries over and can show or delete it. The pattern that works for tutoring, coaching and support calls:

  1. Register a transcript webhook on every call (transcript in the SDK, transcript_webhook on the wire, secret 16 to 200 characters). After the call ends you receive the two-sided transcript, signed. Store the turns, and put a lesson ID in metadata (up to 16 string pairs, echoed on that transcript) to attribute it without a lookup.
  2. Keep who the tutor is in instructions. Keep what happened in context: one system message with a running summary of older material, then the most recent turns.
  3. Stay inside the limits: up to 32 messages, each with the role system, user or assistant and 1 to 8,000 characters of content, and the whole session request under 64 KiB of JSON. Non-ASCII text can use several bytes per character.
  4. Mint the next call with a fresh maxSeconds computed from the remaining balance.
  5. Refresh the summary after each call, so the message window holds the newest turns and the summary holds the rest.
// lib/lesson-context.ts
import type { ContextMessage } from "realtime-avatar";

type Turn = { role: "user" | "assistant"; content: string };

const MAX_MESSAGES = 32; // context limit per call
const MAX_CHARS = 8000; // per message
const size = (messages: ContextMessage[]) => new TextEncoder().encode(JSON.stringify(messages)).length;

// budgetBytes leaves room for instructions and the other fields inside the 64 KiB request limit.
export function buildContext(summary: string | null, turns: Turn[], budgetBytes = 40_000): ContextMessage[] {
  const clip = (text: string) => text.slice(0, MAX_CHARS);
  const head: ContextMessage[] = summary ? [{ role: "system", content: clip(`Summary of the lesson so far: ${summary}`) }] : [];
  const recent: ContextMessage[] = turns
    .filter((turn) => turn.content.trim().length > 0)
    .slice(-(MAX_MESSAGES - head.length))
    .map((turn) => ({ role: turn.role, content: clip(turn.content) }));
  while (recent.length > 0 && size([...head, ...recent]) > budgetBytes) recent.shift(); // drop the oldest first
  return [...head, ...recent];
}

The transcript arrives after the call ends, so a learner who presses Continue at once can beat it; let the mint fall back to the summary you already hold.

Ending the first call cleanly

A call that stops mid-sentence feels broken. call.sayAndEnd(text) speaks one exact line, verbatim, then closes the call, so watch secondsRemaining and hand over before the stop:

"use client";
import { useEffect, useRef, useState } from "react";
import { AvatarCall, createProxyClient, type AvatarCallEndReason, type AvatarCallHandle } from "realtime-avatar/react";

const client = createProxyClient({ proxyUrl: "/api/realtime-avatar" }); // the route above

function WrapUp({ call }: { call: AvatarCallHandle }) {
  const done = useRef(false);
  useEffect(() => {
    if (done.current || call.secondsRemaining === null || call.secondsRemaining > 20) return;
    done.current = true;
    call.sayAndEnd("We're at time for this part. Press Continue and we'll pick up at exercise 4.");
  }, [call, call.secondsRemaining]);
  return null;
}

export function Lesson({ tutorId }: { tutorId: string }) {
  const [part, setPart] = useState(1);
  const [ended, setEnded] = useState<AvatarCallEndReason | null>(null);
  if (ended === "out_of_credits") return <p>Out of minutes for now.</p>;
  if (ended) return <button onClick={() => { setEnded(null); setPart((n) => n + 1); }}>Continue</button>;
  // A new key remounts AvatarCall, so the server mints a fresh call with the replayed context.
  return (
    <AvatarCall key={part} client={client} avatarId={tutorId} onEnded={({ reason }) => setEnded(reason)}>
      {(call) => <WrapUp call={call} />}
    </AvatarCall>
  );
}

When the call ends, the Continue button remounts the call with a new key, your server mints the next call with the replayed context, and the learner sees one lesson with a short pause. Key the end screen off the end reason, as the Calls docs advise: out_of_credits needs a different screen from session_cap, because continuing would fail.

Limits worth knowing

  • 30 minutes is the longest single call Realtime Avatar runs, on every plan. For one unbroken hour or more, pick a plan from the list above with a 60-minute, 2-hour or unlimited limit.
  • Send maxSeconds on every call. The length a call gets is the maxSeconds your server sends, up to 1800, and the docs tell you to compute it from the balance you just admitted; the workspace balance has to cover it.
  • A quiet learner can end a call early through the idle disconnect. call.keepAlive() tells the call the user is still there and postpones it, which helps during long reading or writing exercises.
  • The Sandbox's 8 free minutes cannot hold a 30-minute call. Test the handover with a short maxSeconds, such as 120, before buying minutes.
  • Concurrency is per plan: Starter runs 3 calls at once, Developer 10 and Studio 20, so size the plan for your busiest hour as well as for monthly minutes.
  • Context replay is memory you supply, capped at 32 messages per call. Write what the tutor must not forget, such as the learner's level and the current exercise, into the summary.

Questions developers ask about call limits

Anam stops at 5 or 10 minutes. Which realtime avatar API allows 30-minute calls on an entry-level plan?

Realtime Avatar's $9 Starter runs 30-minute calls with 120 minutes included, then $0.095 a minute. LemonSlice's $8 Starter allows 30-minute calls with 41 minutes included (hosted speech, LLM and voice cost $0.09 a minute extra), Beyond Presence's 49-euro Starter lists unlimited session duration, and Talking Avatar's Voice Agents allow 60 minutes with no monthly fee. Inside Anam, Growth at $299 is the first plan past 10 minutes.

Do I need Tavus Growth at $397 for 25-minute tutoring sessions?

Inside Tavus, yes: Starter stops at 10 minutes and Builder at 15, and Growth is the first plan with a 60-minute limit. Outside Tavus, any paid Realtime Avatar plan runs a 25-minute call. At 100 sessions a month (2,500 minutes), Tavus Growth comes to $769, against $119 on Realtime Avatar Studio or $185.50 on Developer.

Which avatar API allows 30-minute coaching calls without paying $475 for LiveAvatar Business?

Realtime Avatar on any paid plan with maxSeconds set to 1800, LemonSlice Starter ($8), Beyond Presence Starter (49 euros), or Talking Avatar Voice Agents at $0.06 a minute. Thirty minutes is exactly Realtime Avatar's ceiling, so if coaching regularly overruns, use the handover above or a vendor with a 60-minute limit.

What is the longest a single realtime avatar session can run?

On the pages read for this article: 2 to 5 minutes on the free tiers of Anam, Tavus, HeyGen LiveAvatar and Beyond Presence, a 30-minute cap on every Realtime Avatar plan (the Sandbox's 8 free minutes run out first) and on LemonSlice's self-serve plans, 60 minutes on Tavus Growth and Business, LiveAvatar Business and Talking Avatar Voice Agents, 2 hours on Anam Growth and Professional and on LemonSlice Ultra, 3 hours on Talking Avatar's face-only API and 24 hours on LemonSlice Enterprise. Beyond Presence lists unlimited duration on its paid plans.

My live avatar call always ends at 30 minutes. How do I continue the same conversation?

On Realtime Avatar, 1,800 seconds is the ceiling for one call and no plan raises it. Close the call with sayAndEnd, then mint a new one whose context holds a summary plus the latest stored turns, up to 32 messages. The tutor picks up where it stopped because your server told it what happened.

O Anam encerra a chamada em 5 minutos no Starter. Qual API permite chamadas de 30 minutos num plano barato?

O Realtime Avatar limita cada chamada a 1.800 segundos (30 minutos) em todos os planos, inclusive no Starter de US$ 9 por mês com 120 minutos incluídos. O LemonSlice Starter (US$ 8 por mês) também permite chamadas de 30 minutos, e os Voice Agents da Talking Avatar chegam a 60 minutos por US$ 0,06 o minuto. No Anam, o primeiro plano acima de 10 minutos é o Growth, de US$ 299.

Alternatif Anam AI dengan panggilan avatar sampai 30 menit di paket murah untuk aplikasi belajar bahasa

Realtime Avatar membatasi setiap panggilan hingga 1.800 detik (30 menit) di semua paket, termasuk Starter seharga US$9 per bulan dengan 120 menit. Sesi belajar yang lebih panjang bisa dilanjutkan di panggilan baru dengan mengirim ringkasan dan pesan terakhir sebagai context, maksimal 32 pesan. Alternatif lain: LemonSlice Starter (US$8, panggilan 30 menit) dan Talking Avatar Voice Agents (US$0,06 per menit, hingga 60 menit).

リアルタイムアバターの通話が30分で切れます。会話の続きを新しいセッションに引き継ぐ方法は?

Realtime Avatarでは1回の通話の上限は1,800秒(30分)で、どのプランでも同じです。終了前に call.sayAndEnd で区切りの一言を話させ、サーバー側で新しい通話を作成し、保存した文字起こしから要約と直近のやり取りを context として渡します。context は最大32件、各メッセージは1から8,000文字で、リクエスト全体は64 KiB以内に収めます。

Compare and cost guides

Meet the live avatars. Hold the first conversation.

Enter the studio