Image: Atmanity
Atmanity has opened early access to Atmee v2, a real-time avatar and video foundation model aimed at face to face AI agents. Co-founder and CEO Chen-Ping Yu announced the release on November 4 in a long post on X, saying two years of research went into a model that can drive live video calls, voice-only sessions, or short generated clips from one image and about 10 seconds of audio.
After Google acquired my AR startup, I started thinking about the next interface, this time for AI. Text and voice work fine, just like DOS did before the Mac. But nothing’s more intuitive than face to face.
— Chen-Ping Yu (@yuchenping) November 4, 2026
After two years of research, today we’re launching Atmee v2: a… pic.twitter.com/oZWAzGgRH1
Atmanity describes itself as an agentic AI lab founded in October 2024 by a team of AI PhDs whose previous company was acquired by Google. The lab says Atmee v2 renders face, body, micro-expression, emotion, and environment live on a single GPU, with a separate behavior model called Causeur choosing personality, emotion, timing, and listening cues moment to moment. RuntimeWire covered the launch the same day and framed live video agents as stuck between costly general video models and stiff avatar tools.
Yu previously founded Phiar, an AR navigation startup. CB Insights lists Google as acquiring Phiar on September 20, 2022. Yu's X profile still points to Atmanity and an Atmee v2 invite form. In the launch thread he compared text and voice interfaces to DOS before the Mac, arguing that face to face conversation is the more intuitive next step for AI products.
According to Atmee's API docs, every conversation runs in one of three modes. Behavior-driven full-stack has Atmee run behavior, language model, speech recognition, voice, and face. Avatar-only lets a developer bring their own voice agent while Atmee renders the face in sync over LiveKit. Behavior-driven voice-only keeps Causeur and the avatar voice without video. The docs say the model belongs to the avatar, so an Atmee V2 avatar always renders on V2 and a V1 avatar on V1 through the same /v1 endpoints.
We built this live talk show demo to show that live avatars aren’t just for video calls. They can be embedded seamlessly into any application.
— Chen-Ping Yu (@yuchenping) November 4, 2026
From 1:22, the host and the guest are two Atmee v2 avatars talking to each other, live and unscripted. pic.twitter.com/GPSZ9YPBo3
In the same thread Yu posted a talk show demo and said that from 1:22 onward the host and guest are two Atmee v2 avatars talking live and unscripted, meant to show the model inside an app rather than only on a call. Later posts show a DJ inside a music player, artwork that answers questions, tutors and robots in learning apps, 3D characters in games, and a fitness coach avatar. Another clip starts from a Taipei alley image and keeps generating a controllable scene, which Yu called a hint toward an agentic world model while noting distortions and artifacts remain.
Two more mobile demos of Atmee v2 dropped into an app: a personal DJ inside your music player, and artwork you can actually talk to. pic.twitter.com/qgtDtwL9nn
— Chen-Ping Yu (@yuchenping) November 4, 2026
Even a world? This started as a quick experiment: a small alley in Taipei as the starting image, and the same Atmee v2 model generating the scene continuously, live and controllable. There are still distortions and artifacts, but for a live avatar model it hints at where we’re… pic.twitter.com/vRyrDqrQus
— Chen-Ping Yu (@yuchenping) November 4, 2026
Yu also pitched Atmee v2 as a character-focused video foundation model and claimed it can be up to 50 times faster and 10 to 100 times cheaper than general video systems such as Seedance 2.5 and Wan 3.0 for that job. In a follow-up he published a same-prompt comparison claiming Seedance 2.5 took 233.6 seconds and about $0.50 while Atmee v2 finished in 3.5 seconds for about a cent, plus charts placing Atmee near real time at the low end of list prices for short 480p clips. Those figures come from Atmanity. Independent benchmarks were not published with the launch. In replies Yu also wrote that Atmee is not as good as Seedance on quality, but close enough given the speed and cost tradeoff, and that latency can be sub second.
Image: Chen-Ping Yu / Atmanity
Atmee's pricing page lists Free, Starter at $9 a month, Growth at $49, Pro at $199, and Scale at $599, plus custom Enterprise deals. Credits reset monthly. Behavior-driven full-stack on Atmee V2 costs 2 credits a minute, Avatar-only costs 1 credit a minute, and V2 HD rises to 3 and 2 credits a minute. V1 is cheaper and available on every plan. V2 and V2 HD need Growth or higher, or an invite code on Free and Starter. Offline clips start from 0.1 credit. Yu told followers invite codes were being released the week of the launch and pointed people to atmee.ai/v2signup.
Live avatar APIs are already a crowded lane. An October 2026 roundup on realtimeavatar.ai put HeyGen LiveAvatar around $0.09 to $0.18 a minute depending on full or avatar-only mode, Tavus around $0.26 to $0.35 a minute, and Simli as a cheaper render-only face layer. Atmee's Growth plan credit math works out near about $0.27 a minute for V2 full-stack before top-ups, closer to Tavus than to the cheapest render-only options. Exact dollars still depend on discounts and whether Causeur and voice are included. Quality, latency, and concurrency need side by side tests Atmanity has not published against those peers.
Access remains gated. Atmee.ai already hosts Atmee v1 for broader use, while v2 early access is invite-led. Builders who want face to face agents for tutoring, support, coaching, or in-app characters now have another single-image path to try beside HeyGen, Tavus, Simli, and similar stacks. The open questions are whether Atmanity's speed and cost claims hold outside company demos, and whether Causeur's behavior layer is enough to make those faces feel trustworthy in long conversations.


Comments