Renderer or full stack? How to compare realtime avatar APIs without comparing the wrong number

Every realtime avatar API publishes a per-minute rate, and almost every comparison of them puts those rates in one table. That table is close to meaningless, because the products on it are not selling the same thing. Some sell you a renderer: audio goes in, a lip-synced face comes out, and you supply the speech recognition, the language model and the voice. Others sell you the whole conversation and the face is one stage of it.

A renderer at $2.40 an hour and a full stack at $8 an hour are not two prices for one product. They are two products, and the gap between them is a bill you pay to somebody else. So the first question is not what it costs. It is which half of the problem you are buying.

Everything below was checked against each vendor's own pages on 4 September 2026. Where a vendor does not publish a number, this post says so rather than borrowing one from a review site.

The renderers

Simli is the clearest case in the category, and admirably blunt about it. Their SDK page describes it as being for developers who want "full control over the tech stack" and are "building apps with custom voice bots". Their Pipecat guide is the proof: the environment template asks you for a Deepgram key, an OpenAI key and a Cartesia or ElevenLabs key alongside the Simli one. Four vendors, four bills, one conversation.

The most useful thing Simli publishes is a latency budget on their own home page, and it argues our point better than we could. They break a spoken turn into four stages — speech-to-text at roughly 100–500 ms, the language model at 250–450 ms, text-to-speech at 250–1200 ms, and their own speech-to-video at under 300 ms — and add, in their words, "These calculation are estimates. They may not reflect the actual latency of the agent, but it shows where Simli fits in." One slice of four, and they name the vendors you have to assemble for the rest.

bitHuman is a renderer too, but an unusual one: the core product is a portable engine rather than a cloud service. Their architecture page puts it as "audio goes in, lip-synced visual frames come out" at 25 frames per second, from a C++ runtime you can host yourself. That buys the cheapest rendering in the category, and they price the conversation separately when you want it: a managed voice agent is 10 credits a minute on top of the 4 the rendering costs, and a managed agent with vision is 30.

The full stacks

Tavus runs the entire pipeline and documents it in order — perception, conversational flow, speech recognition, the language model, text-to-speech, then the face. They do support a renderer mode, and they attach a warning to it: those modes are "incompatible with Tavus's perception and speech recognition layers", and for the lowest latency they "recommend the full Tavus pipeline".

Anam is the most flexible of the four and says so plainly. Their overview describes a four-stage pipeline and then: "The default setup, which we call Turnkey, runs the whole pipeline for you. You can also bring your own LLM, your own STT, your own TTS, or stream us pre-generated audio and have us run face generation only." Their LiveKit guide is the renderer half of that, described as "Anam handles only the visual avatar. You choose the ears, intelligence, and voice".

Anam also does something almost nobody in this category does: their session analytics endpoint returns per-turn latency percentiles across the transcription-to-first-audio path. If you are evaluating on latency rather than on marketing copy, that is the vendor who has made it possible.

Price per hour of live conversation, like for like

This table compares marginal rates — the published per-minute price once you are past whatever a plan includes — because that is the number that survives a product actually being used, and because for two of these vendors it is the only figure published as plain text. The last column is the part most tables leave out.

Two things worth reading off that. Among the products that run the whole conversation, the spread is roughly four to one from top to bottom — the same job, at $4.80 an hour or at $21. And the two cheapest rows are cheap because they are doing less; add a language model and a voice to the bitHuman rendering line and it lands at $8.40, which is exactly where their own managed agent is priced.

One more thing the marketing does not surface: Tavus bills "a 30-second minimum charge, rounded to the nearest 6 seconds" per conversation, and Anam meters "for the entire duration of the session… regardless of whether you are actively speaking or not", with unused minutes that do not roll over. bitHuman is explicit that idle counts too — a connected avatar looping its idle animation "is rendering, and accrues". If your product keeps a character on screen between questions, read those three sentences again before you model your costs.

Making a character

All four now create an avatar from a single photograph. What differs is how long you wait, and whether you pay per character or get a fixed number of slots.

• Anam: under two minutes, and avatars are slots rather than purchases — one on the free plan up to ten on Professional.
• Tavus: a preview "usable in about a minute" from a photo, then tuning in the background. Custom face slots are allocated by tier.
• bitHuman: the most explicit per-avatar pricing anywhere in the category — 250 credits for a first-generation model, 500 for Essence 2, 2000 for Expression 2 — but the slowest, at roughly 45 minutes to an hour and a half for the second-generation models.
• Simli: a single image, and their API reference says the wait "can take a couple of hours".

Where we fit, including where we do not

Realtime Avatar is a full stack. Speech recognition, the language model, the voice and the audio-clocked rendering are one meter and one bill, and the number in the table above is the whole number. The thing we do differently from the other two full stacks is where your code runs: your tools execute in your page or on your server, with your credentials, and the platform never holds them. The session is minted by an endpoint you own, so who may call and what the character knows are your decisions rather than settings in someone's dashboard.

And the honest limitation, since this post has spent nine paragraphs demanding honesty from other people's marketing: we do not have a renderer mode. If you have already built a voice agent on Pipecat or on LiveKit Agents and you want a face bolted onto audio your own pipeline produces, we cannot do that today and Simli, bitHuman and Anam all can. That is a real reason to choose one of them, and you would find out in an afternoon anyway.

How to choose in three questions

1. Do you already own the conversation? If your agent works and you only need a face, buy a renderer and do not pay for a pipeline you will bypass.
2. Is the character on screen while nobody is talking? Then idle metering is a first-order cost, not a footnote, and the three quotes above matter more than the headline rate.
3. Where do your tools have to run? Every full stack calls functions. They differ in whose machine executes them and who holds the credentials, and that is an architecture decision you will not want to revisit later.

Prices in this category move. Tavus was mid-migration between two pricing schemas on the day this was written, with both still present in the page. Re-check before you commit to anything, including ours — the current figures are always on the pricing page, and the sandbox is free if you would rather just try it.