Tavus pricing: what conversational video really costs

Contents

How much does Tavus cost?

T

Tavus starts free and runs from $22 a month (Starter) up to $975 (Business), with a custom-priced Enterprise tier above that. For most teams building a real product, the $59 Builder plan is the sweet spot. The first month of any paid plan is half price, and annual billing takes 20% off.

But the plan fee is the easy part. Tavus does not charge by the video or the seat; it charges by the live conversation minute, and every call carries a 30-second minimum charge. Once you pass your plan’s monthly minutes you pay per-minute overage on top. So the real question is not “which plan” but “how many minutes will my users actually talk,” and this guide is mostly about that math. All figures are current as of July 2026, read live from Tavus’s pricing page.

Plan	Price / mo	Conversation minutes	Custom replicas	Concurrent	Overage
Free	$0	20	0 (stock only)	1	—
Starter	$22 ($11 first mo)	60	1	1	pay-as-you-go
Builder ⭐	$59	175	3	3	~$0.35/min
Growth	$397	1,300	7	10	~$0.31/min
Business	$975	4,000	15	15	~$0.26/min
Enterprise	Custom	Custom	Unlimited	Custom	Scaling discounts

Prices and allowances current as of July 2026, read live from tavus.io/pricing. The “Overage” column is the metered rate past your included minutes; Free and Starter run pay-as-you-go without a published per-minute rate.

How Tavus pricing actually works

Tavus is not a scripted-avatar tool like HeyGen or Synthesia, and its pricing reflects that. It sells conversational video: real-time, two-way AI humans you talk to, built on its own stack of models. Raven-1 handles perception, so the agent can see and react to visual cues; Sparrow-1 handles turn-taking, so it knows when to speak and when to listen; and Phoenix-4 does the real-time face rendering that makes the result look like a person rather than a puppet.

Running all three live, per user, at once, is genuinely expensive compute, which is the reason Tavus meters time rather than selling a flat subscription like a video tool. The thing you pay for is the minutes one of those agents is live in a session, and everything else about the pricing follows from that.

A conversation minute is measured from the moment a call connects to the moment it disconnects, and it bundles the entire pipeline into one number, so there are no separate charges for the language model, the voice, or the video. That is genuinely simpler than rivals who meter tokens, characters, and render time separately: one Tavus minute already includes perception, the language model, text-to-speech, turn-taking, and WebRTC delivery. You are not assembling a bill from parts.

That simplicity is real, but it hides two mechanics that decide your bill.

The first is the 30-second minimum charge. Every conversation is billed for at least half a minute even if it lasts five seconds, and usage rounds up to the nearest six seconds. The second is subtler: a conversation is billed the moment it is created, whether or not a participant ever joins. Tavus auto-closes an empty session after five minutes, and you can pass test_mode = TRUE to avoid charges while developing, but a buggy integration that opens sessions nobody answers will still run up minutes.

One more mechanic sits on top: annual billing takes 20% off and locks your rate for the term, and you can move between plans, up or down, at any time. So the subscription side is flexible and cheap to change; it is the usage side, the minutes, that you actually have to forecast.

What each plan includes

The plans climb on three axes at once: included minutes, custom replicas (your own AI humans, each trained from a single image or a two-minute video), and concurrent sessions. Here is what each tier is really for.

Free ($0) gives you 20 conversation minutes a month, one concurrent session, and access to Tavus’s stock replica library, with whitelabeled API access even at zero cost. What it withholds is custom replicas and watermark removal. It is a true build-and-test tier, enough to wire up an agent and talk to it, not a preview you cannot use. The 20 minutes go fast once you factor the 30-second minimum, so treat it as a proof-of-concept budget, not a runway.

Starter ($22, or $11 the first month) adds a single custom replica, 60 minutes, transcripts, and watermark-free output, still capped at one concurrent session. It suits a solo developer validating an idea, not a live product with real traffic.

Builder ($59) is the one Tavus marks most popular, and rightly so. It brings 175 minutes, three custom replicas, and up to three concurrent sessions, which is enough for a serious prototype or a low-volume production agent. Three concurrent sessions is the quiet upgrade that matters here: it is the first tier where two users can talk to your agent at the same time without one waiting. For most people shopping Tavus, this is the plan.

Growth ($397) is a real jump, not a step. It carries 1,300 minutes, seven replicas, up to 10 concurrent sessions, conversation recordings, and no per-call duration limit, and its overage rate drops to about $0.31 a minute. You move here when you are productionizing rather than testing, and specifically when either your monthly minutes push past roughly 1,150 or you need more than three simultaneous conversations. Below that, it is more plan than most builders use.

Business ($975) carries 4,000 minutes, 15 replicas, 15 concurrent sessions, the lowest overage rate at about $0.26 a minute, and priority support with a 1:1 onboarding call. It is aimed at products running real, sustained volume. Enterprise is custom-priced for unlimited replicas, a fully white-labeled experience (your branding on the whole flow, not Tavus’s), custom concurrency limits, guaranteed SLAs, and SOC 2 / HIPAA compliance, which is the tier regulated industries and large deployments land on.

What your real bill looks like

The included-minutes number is where a plan is cheap or expensive, so map it to your actual usage before you pick. Effective cost per included minute (price divided by minutes, as of July 2026) falls steadily as you climb: Starter is about $0.37 a minute, Builder around $0.34, Growth roughly $0.31, and Business closer to $0.24. Bigger plans are not just bigger buckets, they are cheaper buckets.

That declining rate is why the overage math decides your plan, not the sticker price. Say you run 500 conversation minutes a month. On Builder you pay $59 plus 325 overage minutes at about $0.35, roughly $173 all in. On Growth you pay a flat $397 but have 1,300 minutes included, so you are nowhere near overage. Builder wins at 500 minutes, but as your usage climbs toward 1,150, Builder’s overage catches up to Growth’s flat fee and Growth becomes the cheaper home.

Now run the same math for three real products. A weekend prototype doing 40 minutes of demo calls lives comfortably on Builder’s 175 included minutes at a flat $59. A customer-onboarding agent handling maybe 800 minutes a month is the awkward middle: on Builder it is $59 plus about 625 overage minutes at $0.35, near $278, so Growth’s flat $397 is not yet worth it, but you are one busy month from it being close.

A support bot fielding 5,000 minutes a month clearly belongs on Business: covering that on Growth runs about $1,544 (its $397 plus 3,700 overage minutes at $0.31), where Business handles it for roughly $1,235 ($975 plus 1,000 overage at its cheaper $0.26). The point where Business starts beating Growth on pure cost is around 3,200 minutes; below that, Growth is cheaper and you would only move up for the extra concurrency and replicas.

The practical rule: find the plan whose included minutes cover your realistic monthly volume, then treat overage as a buffer for spikes, not as your normal operating mode. If overage is a routine line on your bill, the next tier up is almost always cheaper. And model the busy month, not the average one, because Tavus never throttles you, so a traffic spike lands straight on the invoice.

How to keep your Tavus bill down

The per-minute model rewards a few habits, and they are worth building in early rather than discovering on your first big invoice.

Lever	What it saves
test_mode = TRUE while building	Every development minute
Close sessions in code on user-exit	Idle minutes up to the 5-minute auto-close
Right-size the plan, let overage flex	Over-provisioning to a bigger tier
Audio-only mode where a face isn’t needed	Discounts (Tavus says these are coming)
Count replica overage, not just minutes	$40–$65 per extra custom replica

Test with test_mode on. Tavus bills a conversation the moment your code creates it, so an afternoon of debugging can quietly spend real minutes. Passing test_mode = TRUE runs the whole creation-and-connection flow without charging, which is the single biggest saving during development.

Close sessions explicitly. Because an idle session bills until it auto-closes at the five-minute mark, ending conversations in code the instant a user leaves stops you paying for dead air. For a product with lots of drop-offs, that alone can shave a real slice off the meter.

Right-size the plan, then let overage flex. Do not buy Growth to feel safe if your real volume fits Builder; the overage exists precisely so you do not have to over-provision. Buy the tier your normal month needs and let pay-as-you-go absorb the occasional spike. Only move up when overage becomes a habit.

Use audio-only where you can. Tavus runs the same pipeline in an audio-only mode as well as full video, and its pricing page says discounts for the audio-only mode are coming. If a given interaction does not need a face on screen, that cheaper mode is a lever worth watching.

Watch replicas, not just minutes. If your product spins up many branded AI humans, the per-replica overage ($65 on Builder, $40 on Growth beyond your included trainings) can rival the minute cost. Count both before you pick a tier.

Tavus pricing vs the scripted-avatar tools

It is worth being clear why Tavus costs what it does, because a lot of confusion comes from comparing it to the wrong products. Tools like HeyGen and Synthesia sell rendered videos: you write a script, they generate a talking-head clip, and you pay per credit or per finished minute of a file you keep. The cost is bounded and predictable because the output is a fixed asset.

Tavus sells live time. There is no finished file, there is a real-time agent holding a two-way conversation, and you pay for every minute it is running, plus the 30-second floor on each call. That is a fundamentally different, usage-metered model, closer to a cloud-compute bill than a video subscription. It is why Tavus can look expensive next to a $29 HeyGen plan and still be the only tool that does the job: they are not selling the same thing. If your use case is scripted video, the per-minute model is simply the wrong meter, and that is the real signal to look elsewhere.

There is also a separate consumer product to avoid confusing with the developer pricing above. Tavus’s “PALs” plans ($0, $20, and $50 a month) are personal AI-companion subscriptions billed in flat call minutes, not the API-metered CVI product a builder wants. The tiers in this guide are the developer plans; the PALs plans are a different offering entirely.

Pitfalls the pricing page soft-pedals

• The 30-second minimum punishes short sessions. A product built on quick, one-question interactions pays for 30 seconds each time, so 200 ten-second chats cost the same as 100 minutes of talk. If your sessions are short and frequent, model the minimum, not the average.
• Empty sessions still bill. Because a conversation is charged on creation, a front-end bug or an abandoned call that never connects can accrue minutes. Use test_mode while developing and close sessions explicitly in production.
• Video generation is a separate meter. Tavus’s async MP4 video generation has its own small minute allowance (5 to 100 by tier) and its own ~$1-a-minute overage. The plan’s big headline number is CVI (live conversation) minutes; do not budget video generation against it.
• Custom replicas have their own overage. Past your plan’s included replica trainings, each extra custom replica is a one-time fee, about $65 on Builder and $40 on Growth. A team spinning up many branded personas should count replicas, not just minutes.
• Concurrency, not just minutes, can force an upgrade. Free and Starter allow only one live session at a time, so the first time two users try to talk to your agent at once, one waits, no matter how many minutes you have left. Any real product needs Builder’s three concurrent sessions or more, which means concurrency, not minute count, often sets your floor.
• Annual billing locks the rate. The 20% annual discount is real, but it commits you for the term. Take it once your volume is stable; on a monthly plan while you are still finding product-market fit, the flexibility to downgrade is worth more than the discount.
• It is still in beta. In our hands-on Tavus review the technology impressed but the first call dropped before reconnecting, so budget for some roughness alongside the bill.

Which Tavus plan should you pick?

• Just exploring the API? Start on Free (20 minutes) and use test_mode so development does not eat them.
• A solo developer validating one idea? Starter ($22) buys a custom replica and 60 minutes; enough to demo, not to launch.
• Building a real product prototype or a low-volume agent? Builder ($59) is the default answer, with three replicas, three concurrent sessions, and 175 minutes.
• Running a production agent past ~1,150 minutes a month or needing 10 concurrent sessions? Move to Growth ($397) before overage on Builder passes its flat fee.
• High volume, compliance, or white-labeling? Business ($975) or Enterprise. Skip these unless you genuinely need 15+ concurrent sessions, SOC 2 / HIPAA, or your own branding on the whole experience.

The one number that overrides all of these is concurrency. You can be well within your minutes and still need to move up because your product briefly has more simultaneous conversations than your tier allows, and that ceiling, not the minute count, is what makes an agent feel broken to a user who has to wait. Size the plan for your busiest concurrent moment first, then check the minutes.

If Tavus isn’t the right fit

Tavus is priced and built for real-time conversation, so the honest alternative depends on whether you actually need that. If you only need an avatar to read a script rather than hold a live conversation, you are paying for the wrong thing, and a scripted-avatar tool is cheaper and more mature.

See our HeyGen pricing guide for the credit-based option and our Synthesia review for the enterprise-training one; both make talking-head video far more cheaply than Tavus’s per-minute live pricing, and our best AI avatar generator roundup lines up the scripted-avatar field by use case. If you landed here after weighing Synthesia, our Synthesia alternatives guide covers where Tavus fits among them as the real-time pick.

But if you genuinely need a two-way AI human that listens and responds in real time, there is no cheap shortcut, because that is the category Tavus effectively defines. Most tools that come up as Tavus alternatives target marketing-led personalized video rather than the developer-grade real-time conversational API Tavus sells, so they solve a related but different problem. In that case the decision is which Tavus tier, not which rival, and the plan-by-plan guide above is the whole answer.

Final word

Tavus pricing looks simple, six tiers from free to custom, and it mostly is. The complexity that catches people is under the headline number: you pay by the live conversation minute, every call bills at least 30 seconds, and overage past your allowance is where the real cost lives. Pick the tier whose included minutes match your realistic volume, and the bill stays predictable.