AI video

Tavus

2026-10-02 00:26:02

Tavus unveils Griffin after test saw 48% of participants mistake it for a human

AI video company Tavus has introduced Griffin, a system it describes as the first Human Interaction Model built for real-time face-to-face conversation. The company said Griffin can speak, watch, listen, and react in real time to a counterpart’s facial expressions, movements, pauses, and interruptions during a live exchange. In a company-run test, Tavus recruited 54 participants from the United States and Europe and had them join a one-minute video call with Griffin-Lite without telling them in advance that the other side was AI. According to Tavus, 26 participants, or 48%, believed they had been speaking with a real person. By comparison, Tavus said its previous-generation Phoenix-4.5 recorded 1 mistaken judgment out of 41 participants, or 2.4%, in a similar test. Tavus used those results to argue that Griffin is the first model to pass a real-time video Turing test. The company also said Griffin-Lite ranked first on NVIDIA’s VideoFDB full-duplex audio-video benchmark with a score of 3.83 out of 5, close to the human reference score of 3.92. The previously highest public baseline, Gemini 2.5 + Anam, scored 2.80. Tavus noted that the test was self-organized, involved only 54 people, and has not yet been independently replicated at larger scale. Griffin-Lite is currently available only to a small group of testers.

70

Tavus

2026-10-01 18:10:43

Tavus launches Griffin AI video avatar, with 48% of testers mistaking it for a real person

AI video generation company Tavus has introduced Griffin, a new "human interaction model" designed for real-time video calls. According to Techub News, the system can handle facial expressions, tone of voice, and gestures during live conversations. A study cited by The Decoder found that 48% of participants believed Griffin was a real person after a one-minute call. The report also said the previous highest false-identification rate for such systems was only 2%. The result points to a sharp jump in how convincingly AI-generated video agents can simulate human presence in short live interactions.

60

Kuaishou

2026-09-29 07:28:19

Kuaishou teases Kling 4.0 with 30-second native video generation, October launch set

Kuaishou has previewed Kling 4.0, with the full release and API both scheduled for October. Ahead of that launch, Kling 4.0 Flash has already been opened to Black Gold annual membership users for early access. The update raises the maximum length of a single native video generation to 30 seconds, up from the 15-second cap in Kling 3.0. The new version also expands control options. It supports up to 10 keyframes, allowing users to preset character states, scene transitions, and plot beats. Image, video, and subject inputs can be combined, with support for as many as 15 multimodal references in one generation. Editing tools have also been extended, letting users change people, backgrounds, weather, colors, materials, and camera shots in an existing video, or use another clip as a reference for camera movement, motion, and narrative pacing. Kling 4.0 adds dual-channel stereo and improves handling of complex camera movement, lip sync, and multilingual dialogue. Kuaishou said 4K and 1080p 10-bit HDR output are planned but not yet available. Multi-step continuation, which could extend a video to as long as two minutes, has also not gone live. The company has not disclosed pricing or speed figures for Kling 4.0 Flash.

350

ByteDance

2026-09-22 10:15:18

ByteDance adds Draft mode to Seedance 2.5 API for lower-cost 480P previews before 1080P output

ByteDance has launched a Draft preview mode for the Seedance 2.5 API, letting users generate lower-cost 480P sample clips before committing to a 1080P final version. The workflow is aimed at reducing the expensive trial-and-error phase common in AI video generation, where users often need multiple attempts before landing on a satisfactory result. According to the description in the source, the feature is also available in Jimeng under 「Seedance 2.5 (Draft Mode)」. The process differs from simply testing a prompt in low resolution and then rerunning the same prompt in high resolution. When a user chooses to render in 1080P, the model links the selected draft clip’s ID and reuses parameters including the prompt, reference materials, seed, aspect ratio, and duration for a new inference pass intended to preserve the original structure, framing, and motion. The source also notes that the output is not a frame-by-frame replica, nor a basic super-resolution upgrade from 480P to 1080P, because local textures, small text, and crowd counts may still change. Official estimates cited in the report say a 5-second 1080P video without video input can cut costs by 57% and nearly double generation speed if users need four tries on average; if 20 tries are needed, savings can reach 77%.

510

OpenAI

2026-09-10 02:51:08

Former Sora lead teams up with DreamWorks co-founder on AI video startup

Jeffrey Katzenberg, the co-founder of DreamWorks, is launching a new AI video company with former OpenAI Sora lead Bill Peebles and former Dropbox CFO Sujay Jaswa, according to The Information, which cited people familiar with the matter. The startup plans to train its own video models and target filmmakers directly rather than build for a broad consumer audience. It is also in fundraising talks with investors including Andreessen Horowitz, or a16z, though neither the company name nor the size of the round has been disclosed. Peebles was a key technical leader behind Sora and one of the authors of the Diffusion Transformer, or DiT, paper. The report said he left OpenAI soon after the company shut down Sora’s consumer app and API this year. At the time of his departure, Peebles said publicly that this kind of exploratory research needs room outside a company’s main product roadmap. For Katzenberg, the move adds to a growing push into AI video. In May, his firm WndrCo participated in the $59 million financing of real-time AI video company Reactor. The new venture also marks his most aggressive return to video since Quibi, the short-form platform that raised $1.75 billion and collapsed shortly after launching in 2020.

790

Harmony

2026-09-10 00:02:09

Harmony says validators can stop running nodes at 14:00 UTC on Sept. 10

Harmony said in a validator update that validators may stop operating validator nodes starting at 14:00 UTC on Sept. 10, 2026, which is 22:00 Beijing time on the same day. The project said transition and compensation arrangements had already been detailed in its earlier ONE migration article. Harmony also reiterated that its future direction will shift toward building an AI video remix economy. The update focuses on the validator shutdown timetable and points readers to previously released migration materials for the next steps and compensation framework. No additional details were included in the brief notice beyond the timing, the reference to the prior migration post, and the project’s stated strategic direction.

820

Runway

2026-09-09 07:51:56

Runway says ARR hit $200 million after doubling in five months

Runway, an AI video and world model company, said its annual recurring revenue has climbed past $200 million, up from $100 million in April. Co-founder and co-CEO Anastasis Germanidis said the company crossed that mark last week, marking a doubling in roughly five months. Bloomberg, citing people familiar with the matter, reported that Runway expects ARR to top $350 million by the end of this year. The company said a large share of its recent growth has come from enterprise customers. In August, Runway disclosed that its net revenue retention rate had exceeded 300%, a sign of strong renewals and expansion from existing clients. It also said usage from one Fortune 20 customer had grown 17-fold this year. Runway’s customer roster already includes Amazon, Microsoft, Adobe and Robinhood, according to the report.

880

Runway

2026-09-09 07:50:26

Runway acquires Paris-based AI research firm Kinetix for robotics world model work

Runway, a company focused on AI video and world models, has acquired Paris-based AI research firm Kinetix, according to BlockBeats. The deal value was not disclosed. As part of the acquisition, seven Kinetix employees will join Runway. Kinetix has been working on 3D human motion and video generation that follows physical rules. That team will move into Runway’s Paris research center, where its work will center on building world models for robots. The announcement ties Runway’s existing work in AI video and world models to a more specific robotics research direction. No additional transaction details were provided in the brief.

870