AI video's new race is real time: Google cuts prices, Runway talks back live
Google cut its video model's price 30% and made 15-second clips editable in conversation. Runway made a 57-billion-parameter model talk back in real time.
- Gemini Omni 1.2 Flash makes native video up to 15 seconds, costs 30% less than its predecessor, and is priced at $0.07 per second for 720p video. Self-reported m:…9065
- Omni 1.2 Flash adds multi-turn conversational editing across full 15-second clips, with durations selectable in one-second steps from 3 to 15 seconds. Self-reported m:…9065
- Runway's Characters system runs a 57-billion-parameter video diffusion transformer at 24 frames per second, averaging above 23.5 fps in 99.5% of sessions over seven weeks. Self-reported m:…3840
- Runway cut per-step time from 240.9ms at launch to 147.4ms, including cache upkeep falling from 69.7ms to 6.2ms per step. Self-reported m:…3840
- Felo now offers four video models — MiniMax H3, Seedance 2.5, Wan 3.0 Prime, and Kling 4.0 — of which only Seedance 2.5 and Wan 3.0 Prime generate sound. Self-reported m:…7147
What changed
Google DeepMind has launched Gemini Omni 1.2 Flash, a video model that supports native video creation up to 15 seconds, at a price 30% lower than its predecessor, Gemini Omni 1.1 Flash. It is priced at $0.07 per second for 720p video and is generally available on the Gemini API and Google Cloud. Users can specify video duration in one-second increments from 3 to 15 seconds, and outputs carry an imperceptible SynthID watermark. m:…9065
The more interesting change is how the model is directed. Omni 1.2 Flash introduces multi-turn conversational editing across full 15-second clips, and supports reference adherence using character, product, or style images, short motion clips of up to three seconds, and audio cues. That turns video generation from a one-shot prompt into a conversation — change the character, keep the scene, adjust the motion — which is how creators actually work when a client asks for a revision. m:…9065
Google is also giving the model away at the low end. Google Vids now runs on Omni 1.2 Flash, and Google account and Workspace users can generate up to three 15-second videos per month in Vids at no cost, with improved speech narration for social clips and explainers. A free tier of three short clips a month will not replace a production workflow, but it puts the editing loop in front of every Workspace user, which is how a tool becomes the default. m:…5548
Real time is the harder problem
Runway attacked a harder problem: a video character that talks back. Its Characters system uses a 57-billion-parameter autoregressive video diffusion transformer, and real-time conversation needs 24 frames per second, meaning a new four-frame chunk every 166.7 milliseconds. The initial launch used eight H100 graphics cards per session and took 240.9 milliseconds per diffusion step — too slow to hold a conversation. m:…3840
The fix was engineering, not a new model. Optimizations cut step time to 159.3 milliseconds, and a fused communication kernel cut it further to 147.4 milliseconds. Cache upkeep fell from 69.7 milliseconds to 6.2 milliseconds per step, and fusing tensor-parallel all-reduce into FP8 matrix multiplication hides 54 to 68% of its time. Decoder offloading to a separate graphics card adds only 1.6 milliseconds to the step period. The result averaged above 23.5 frames per second in 99.5% of sessions over seven weeks, with median turn latency from user speech to character voice of 2.13 seconds. m:…3840
The numbers that matter for a product team are the latency ones. Pipeline latency from agent audio to speaking frame is 0.49 seconds, and production now runs on three B200 graphics cards at 154.3 milliseconds per step. A video character that answers in about two seconds is usable; one that answers in five is a demo. Runway's post is effectively a recipe for the second kind of video product — not the clip you render and post, but the character you talk to. m:…3840
What it means for creators
The market is fragmenting fast enough that comparison guides are now a product of their own. Felo's guide to its four integrated models — MiniMax H3, Seedance 2.5, Wan 3.0 Prime, and Kling 4.0 — shows the trade-offs: MiniMax H3 runs 4 to 30 seconds at lower resolutions and 4 to 15 seconds at 1080p or 2K, Seedance 2.5 runs 4 to 30 seconds, Wan 3.0 Prime runs 2 to 30 seconds, and Kling 4.0 runs 3 to 30 seconds. Only Seedance 2.5 and Wan 3.0 Prime generate sound, and all four support 16:9, 1:1, and 9:16 aspect ratios. Choosing a model is now a spec-sheet exercise, not a taste test. m:…7147
Both vendors are honest about the limits. Google says development is ongoing to improve fast sports action, rapid script changes, and complex physics in Omni 1.2 Flash. Runway reports mean lip-sync differences of at most 0.03 across 51 scored clips — good, not perfect. The honest read is that short, controlled clips with a consistent character are now cheap and fast, while chaotic action still belongs to a camera. Creators should price their ambitions accordingly: the 15-second product promo is solved, the action sequence is not. m:…9065 m:…3840
What is still unresolved
- Whether conversational editing in Omni 1.2 Flash holds character consistency across many turns, not just the first few.
- How Runway's real-time character economics work at scale, given production still needs three B200 graphics cards per session.
- Which of the four Felo models creators actually keep using once the free and trial credits run out.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.