Meta's New Image Model Writes Code Before It Draws — and the Real Buyer Is Advertisers

Meta Superintelligence Labs’ new image model doesn’t map a prompt straight to pixels — it works agentically, invoking search and coding tools mid-generation, self-refining its own output, and improving with more test-time compute. Muse Image can write and execute code to produce accurate plots and QR codes, pull in personal photos through an @-mention feature, and ground images in real-time web references rather than training-data guesses alone. A companion model, Muse Video, previewed alongside it adds native audio on the same pretraining base. Muse Image is live now in the Meta AI app, on Instagram Stories in the US, and on WhatsApp in limited countries — and it will power Advantage Plus, Meta’s automated ad-creative tool for advertisers, which is the part of the launch that matters more than the consumer rollout.

The launch landed a week after Google DeepMind cut its own image-model prices, releasing Nano Banana 2 Lite at $0.034 per image and Gemini Omni Flash at $0.10 per second of video. Between the two moves, image and video generation are getting more capable and cheaper in the same month: Meta’s agentic self-refinement raises the ceiling on what a generated image can do correctly, while Google’s pricing drops the floor on what it costs to generate one. For any company weighing whether to build creative production in-house versus keep commissioning it out, both curves are moving in the same direction at once.

Muse Image also drew immediate pushback from talent agency CAA over its opt-out — rather than opt-in — handling of user photos. That’s the detail worth remembering before treating any of this as settled: the technical capability is arriving faster than the rights and consent framework around it, and a company adopting these tools for its own marketing needs a policy for image sourcing before a vendor’s default settings become its own liability.