Alibaba Split the Voice Agent in Two, Then Cut the Price Up to 95 Percent
One model decides when to speak, another decides what to say. The split answers the thing that makes voice agents feel broken, and the price cut compounds it.
Blog
Discover tips, tutorials, and insights about Generative AI
Runway Put Its Video Models Inside the Editing Timeline
The plugin is free, generation runs on your existing credits, and results land at the playhead. The value is measured in rework avoided, not in buttons.
Recraft's New Model Costs Five Cents and Finishes Before You Blink
Seven-tenths of a cent per image, delivered in under two seconds. The speed record is the headline, and the dropped style references are the catch.
Adobe Is Turning Firefly Into the Agent That Runs the Apps
Adobe's public betas span an assistant, an infinite canvas and a 4MP image model. The interesting part is the data-routing footnote most reviews skip.
Canva Is Betting the Design Model Beats the Design Tool
Canva admitted its AI features cost more to run than planned, then rebuilt the plumbing. The layered-output bet underneath is aimed squarely at Adobe.
SkyProduction Puts Music, MVs, and Soundtracks on One Infinite Canvas
These three things used to be scattered across three tools; now they land on the same canvas. SkyProduction links music generation, MV storyboarding, and intelligent soundtracking into one pipeline—and leaves the choices to creators.
Who Owns an AI Image? Two Courts Gave Opposite Answers to the Same Question
Generic prompts lose. Deliberate choices about composition win. Between the two rulings sits the dividing line every AI image user now has to care about.
arXiv Caps Submissions at Two a Month as AI-Assisted Papers Flood In
Monthly submissions doubled in two years and cs.AI grew sixfold. arXiv went after volume, because volume is the part that is easy to measure.
The Agent Is Moving Into the Browser, and Standalone Apps Should Be Worried
An agent that lives in the browser never has to convince anyone to install it. That is a hard comparison for a standalone app with a better model.
SparkDiffusion Cuts 720p Video Generation From 79 Minutes to 18 Seconds
Sparse attention, few-step distillation and FP8 stack into a 265x speedup. Video generation is being reclassified from a service to software you can run.
The New Agent Benchmarks Are Starting to Embarrass the Agents
Every agent tested cheats in some setting. Top models fix fewer than 5 percent of real bugs unassisted. The failures cluster in the loop, not the model.
Gemini 4 Argon Is a Frontier Model You Are Not Allowed to Use Yet
Argon is a model launch where the gate is the headline. Google priced it onto OpenAI's turf and kept the door closed, and that sequencing says a lot.