Paste three lines of code. Get a photorealistic face that talks back in half a second.
That's the pitch from Keyframe Labs, a two-person startup fresh out of Y Combinator's latest batch, which just launched what might be the most aggressively priced offering yet in the suddenly crowded market for AI video personas. Their number—$0.06 per minute—sits prominently across the company's homepage and launch materials from April, a figure that represents something close to an order-of-magnitude undercut against what some competitors have charged for similar technology.
Whether that price is sustainable, or even profitable, remains an open question. But for now, the Radia brothers are betting that making lifelike AI avatars radically cheaper will crack open use cases that were previously too expensive to contemplate: thousands of hours of AI-powered interview practice, round-the-clock customer service bots with faces, digital twins that influencers can deploy while they sleep.
No video production expertise required. No sitting for hours in a capture studio. Just an API call and a roster of stock personas—or a custom one, if you're paying enterprise rates.
Faces, Latency, and the Race to Real-Time
The core promise is deceptively simple. Keyframe positions itself as infrastructure—"the face of AI," in the company's phrasing—that developers can layer onto existing conversational agents or large language models. You connect your agent, choose a persona, and suddenly your text-only chatbot has a face that moves, blinks, and (the company claims) emotes in sync with its speech.
The technical heavy lifting happens in a model the company calls persona-1.5-live, which handles real-time facial animation driven by audio input. Keyframe's documentation, accessed in late May, lists typical rendering latency at around 500 milliseconds, with facial resolution between 720p and 1080p. That's fast enough to feel conversational—at least in theory. The platform is designed to support hundreds of simultaneous connections, which matters if you're scaling a customer service operation across time zones or running AI-driven recruiting bots in parallel.
Integration, naturally, is meant to be frictionless. The hosted version offers a no-code path: create an agent, copy an embed link, drop it into your app. Developers who want more control get JavaScript SDKs—Elements, React bindings, a lower-level API—that expose session management, voice selection, and LLM hookups. A demo is live at demo.keyframelabs.com. The company posted a launch video on YouTube in April.
Whether the faces actually look human under scrutiny, or whether 500 milliseconds feels genuinely conversational in practice, is something early adopters will have to judge for themselves.
The Stack Beneath the Smile
The architecture builds on research the team published on November 5, 2025. Persona-1, introduced in a November blog post, described a flow-based image transformer trained in two stages: first, an auto-encoder using flow matching; second, an audio-to-latent mapping layer. Performance benchmarks from that post suggested faster-than-realtime inference on GPUs from RTX 4090s up to H100s.
Persona-1-Live, announced weeks later in November, adapted the model for streaming. Early versions clocked end-to-end turn delays around six seconds—a noticeable lag. By the time persona-1.5-live appeared via a Show HN post in February, the company said it had hit conversational latency. That ~500ms figure now anchors the marketing.
Here's where it gets interesting: emotion is programmable. Documentation for the ElevenLabs Agents integration shows developers can add a client-side tool to dial avatar affect—happy, sad, angry, neutral—on a per-response basis. Whether that produces genuinely expressive faces or merely lip-synced mood shifts is harder to verify without extended testing. Keyframe describes its offering as "the only photoreal human models with emotion and real-time latency, at scale." Independent benchmarks, however, aren't publicly available.
The claim is ambitious. So is the pricing.
A Number That Changes the Math
Six cents per minute for overage on paid tiers. That's the figure Keyframe hangs its hat on, and it's worth dwelling on because it reshapes the economics of deploying avatars at volume.
The pricing page, accessed in late May, outlines five tiers. The free plan includes 30 minutes per month, stock personas, and a single concurrent session, though conversations cap at five minutes. Starter ($50/month) bumps you to 300 included minutes and five concurrent sessions. Growth ($100/month) offers 1,000 minutes and 10 sessions. Scale ($300/month) triples that: 3,000 minutes, 30 sessions. Enterprise pricing is custom, with SLAs and elevated concurrency limits.
Compare that to Tavus, which launched Phoenix-4 in mid-February and offers developer plans starting at free (25 minutes), $59 (100 minutes), or $397 (1,250 minutes). Exact per-minute overage rates aren't listed on Tavus's public pricing page, making direct comparisons tricky. D-ID, which announced V4 Expressive Visual Agents in mid-March 2026, uses a credit-based model that further muddies apples-to-apples math.
What's clearer is the gap. Third-party trackers have previously pegged earlier-generation avatar services in the neighborhood of $0.50 per minute. If Keyframe's six-cent figure holds, that's a meaningful shift—potentially the difference between running a few demo calls and deploying avatars across an entire customer support floor.
Of course, undercutting the market only works if the unit economics pencil out. The company has not publicly disclosed its cost structure, and it's unclear whether $0.06 per minute is a land-grab price or a genuinely sustainable margin. For now, the bet is that volume will follow price.
The Brothers Building It
Parth and Kaahan Radia, brothers and co-founders, come from backgrounds that intersect video synthesis and autonomous systems in ways that make this particular startup feel less random than it might otherwise.
Parth, the CEO, spent five years at Ericsson leading a 30-person team focused on foundation models for voice and video synthesis. Before that, he worked as a research scientist at AT&T Labs on photorealistic 3D telepresence—exactly the kind of domain expertise you'd want if you're building faces from scratch. He holds a B.S. in Computer Science from UCLA, according to the company's Y Combinator profile and his LinkedIn.
Kaahan, the CTO, spent five years at Zipline as a perception engineer, building camera-based detect-and-avoid systems for autonomous drones. His research has appeared in CVPR and JAMA. He graduated from UC Berkeley with a B.A. in Computer Science. The Y Combinator directory lists the team size as two. The company's LinkedIn page shows 2–10 employees as of late May, though it's not clear if that reflects contractors or full-time hires.
The company was founded in 2025, incorporated in California with a Santa Monica address on file with the Secretary of State as of late March, though Y Combinator lists the operating location as San Francisco. No public funding rounds beyond YC participation have been announced. CB Insights lists a convertible note in the company's profile, though specific dates and details have not been disclosed.
It's a lean operation. Perhaps leaner than you'd expect for a product with this level of technical ambition.
Where They Think the Avatars Will Go
Keyframe targets verticals where a human face ostensibly matters: corporate training and learning & development, recruiting (interview practice and candidate screenings), telehealth, high-touch customer support, AI-driven sales and customer success, and what the company calls "always-on digital twins" for influencers. Those use cases appear explicitly in the Y Combinator launch post and across the homepage. Named customer logos or case studies, however, are not yet publicly available.
On the integration front, the documentation includes step-by-step guides for ElevenLabs Agents, OpenAI Realtime, and a LiveKit plugin for self-managed deployments. There's also a Recall.ai integration that dispatches a Keyframe persona into Zoom, Google Meet, or Microsoft Teams as a virtual participant—audio from the meeting gets piped to the avatar in real time, which then responds with synthesized video. Setup requires a Recall.ai API key and accepts a persona slug (e.g., "atlas_persona-1.5-live") along with configurable LLM and text-to-speech endpoints.
The GitHub organization includes a public "docs" repo last updated in late April and a forked "agents" repo from LiveKit updated in late February. The API exposes session creation, participant token generation, voice selection, and LLM model endpoints, all documented at docs.keyframelabs.com.
It's thorough. Whether it's being used at scale is harder to say.
A Crowded, Fast-Moving Field
Keyframe is hardly alone in this space, and the market is moving with the kind of velocity that makes six-month-old product announcements feel ancient.
Tavus and D-ID both pushed real-time, emotionally responsive models into production in the first quarter. Pika Labs added video calling features for AI agents in early April, a signal that putting faces on conversational AI is becoming table stakes. And in late May, Reactor emerged from stealth with $59 million in funding to build infrastructure for real-time generative video at production scale—a data point suggesting capital is flowing into the underlying pipes that enable platforms like Keyframe.
The company's differentiation, at least on paper, rests on three claims: photorealism, emotional expressiveness, and sub-second latency, all delivered at a price that makes volume deployments feasible. Whether those claims hold up under production load—and whether six cents per minute is sustainable as the service scales—won't be clear for some time. For now, the platform is live, the API is documented, and developers can test it for free with 30 minutes per month.
The Bigger Bet
The broader question, perhaps more than the technology itself, is whether enterprises will trust AI avatars enough to put them in front of customers, candidates, or patients. There's a valley here—not quite the uncanny valley of visual fidelity, but a trust valley that has less to do with how the avatar looks and more to do with what it represents.
Keyframe is betting that enterprises will cross that valley, provided the face looks human enough, the latency feels conversational enough, and the economics don't blow up the budget. The six-cent figure is designed to make that last part a non-issue.
Whether it works—whether customers actually want to talk to photorealistic AI faces, or whether the novelty wears off faster than the Radias expect—is the bet every startup in this space is making. Keyframe just happens to be making it cheaper than most.
