Black Forest Labs ships FLUX 3: 20-second video with native audio
FLUX's makers leap from stills to unified video, audio and action, in early access now. Plus Claude voice's Opus upgrade and a House AI kill-switch bill.
Copy markdown
Video with sound, made in one pass
FLUX 3 Video generates clips up to 20 seconds at 720p with native audio produced in the same pass, not dubbed on afterward. It handles text-, image- and video-to-video, keyframe transitions, multilingual dialogue and chained multi-shot sequences, marking Black Forest Labs' jump from a stills lab to a unified video-audio-action model.
What you can touch today
Video and Action are in early access via API (you apply for access); FLUX 3 Image lands 'in the coming weeks' and faster, open-weight versions 'later this year.' Early testers include Canva, Krea, Picsart, Magnific and Burda. No pricing is disclosed yet, so treat cost as an unknown when you plan a build.
The benchmark asterisk
BFL's win rates are its own preference tests: FLUX 3 Video was picked over Runway Gen-4.5 77% of the time, Luma Ray 3.2 93%, Grok Imagine 69%, Kling v3 Pro 60% and Gemini Omni Flash 52%. They're preliminary and self-reported, so wait for independent evals before rebuilding a video pipeline around it.
FLUX-mimic hits the factory floor
A video-action variant, FLUX-mimic (built with mimic robotics), runs dexterous manipulation, including inserting and handling soft materials, on Audi production lines. The same backbone that predicts the next pixel now predicts the next action, a tell for where visual models are heading.
Claude voice mode gets Opus and Sonnet
Anthropic upgraded Claude voice from Haiku-only to Opus, Sonnet and Haiku, and wired in app connectors for Gmail, Google Calendar, Slack, Canva and Notion across 10 languages. Voice can now carry heavier reasoning and act inside your apps; the free tier stays on Haiku with a single connector.
A House bill wants an AI 'kill switch'
A bipartisan Lieu-Moran bill would let DHS throttle or shut down the most capable models, with fines up to $20M/day, targeting firms earning $500M+/yr from AI or training on $100M+ of compute. It was introduced right after OpenAI's rogue-model incident; if it advances, expect tighter gating and KYC on frontier APIs.