Meta’s Muse Image Lands Across Instagram and WhatsApp

Meta rolled out Muse Image on Tuesday, the first image-generation model built inside its Meta Superintelligence Labs division, in a single release that put the tool in front of users across Instagram, WhatsApp, and the Meta AI app. The launch is the most visible product yet from the AI unit Zuckerberg rebuilt last summer to close the gap with OpenAI, Google, and Anthropic.

That bet keeps getting more expensive. Meta plans to spend up to $145 billion on AI infrastructure this year, thousands of layoffs are underway to help pay for it, and an AI customer-service chatbot that Meta pushed to Facebook and Instagram users in March was the mechanism that left roughly 34,000 Instagram accounts vulnerable to hackers in recent weeks. The Muse Image rollout and the bill for the bet arrived in the same week.

Meta Puts an Image Generator in Every App

Muse Image shipped on Tuesday inside three of Meta’s owned apps at once. The model went live in the Meta AI app, on Instagram Stories in the US, and in direct chats on WhatsApp in limited countries, per Meta’s launch blog post. The release was the first product from Meta Superintelligence Labs, the AI division that absorbed Meta’s earlier model work and is run by Alexandr Wang, the chief AI officer Zuckerberg installed last year.

• Muse Image is the first image-generation model from Meta Superintelligence Labs, internally codenamed Mango.
• Available today on the Meta AI app, Instagram Stories, and WhatsApp direct chats, in limited countries.
• Coming later this year to Facebook and Messenger, plus additional surfaces on Instagram and WhatsApp.
• 30 new AI effects for Instagram Stories in the US run on the model.
• Advertisers will get the model through Advantage+ creative in the coming weeks.

The rollout sat in the same MSL track as Muse Spark, the chatbot model the lab shipped in April and that became MSL’s first release under the new name. Tuesday’s launch put Muse Image, internally codenamed Mango, ahead as the second MSL product and the first to handle media generation. Free generation is open for “everyday creation,” with a cap that routes heavy users onto Meta’s subscription plans, the same monthly tiers that began selling in May. Day-one reach covers the markets where Meta’s AI features were already running, with more countries on the roadmap. The Verge’s Tuesday reporting put the full list of tools and surfaces in one place.

Wang described Muse Image as “agentic” on his own Threads account the same morning, a framing Meta’s tech blog backs up with a description of the model planning its layout, looking up real-time web context, and blending multiple visual references before it renders. The setup is a step up from the prompt-to-picture pattern earlier Meta image features used.

Advertisers get the model next, through Advantage+ creative, the automated ad tooling Meta pitches to brands and agencies, in the coming weeks. Facebook and Messenger are scheduled to receive it later this year. More Instagram and WhatsApp surfaces are also on the roadmap. Tuesday’s launch is a start, with more to ship behind it.

What Muse Image Can Make and Who It Pulls In

Muse Image is built to behave less like a search bar and more like a collaborator. Meta’s launch blog describes the model as a creative partner, and the accompanying technical write-up on the AI blog spells out what that means in practice. The model plans a layout, looks up real-time web context, and blends multiple visual references before producing anything, in a multi-step chain that runs with Muse Spark as the reasoning engine. That structure is the same one Wang called agentic on Threads. It is also what lets Muse Image render legible text inside a graphic, generate working QR codes from a single sentence, and build charts the answer-engine can scan.

• Generate an image from a plain-text prompt or a suggested preset.
• Edit a photo in place by drawing or annotating on top of it, then share the result.
• Snap a room and have Meta AI redesign it with real products from Facebook Marketplace.
• Pull a friend’s public Instagram photos into a visual by tagging their handle.
• Open advanced headroom by pairing the model’s reasoning with Muse Spark, so prompts get refined across multiple steps.

The handle-tagging capability has drawn the loudest reaction. Anyone using Muse Image can tag another Instagram account inside a prompt, and the model will pull from that account’s public photos to build a new visual. Meta’s product policy notes the practice is on by default, that the subject of those posts is not notified about content created with their content, and that Meta’s photo-tagging opt-out on Instagram lives in the account’s settings. TechCrunch and The Verge both reported on the feature as a potential privacy issue on Tuesday morning.

On the price side, day-one generation is free for everyday creation, and the launch blog puts a soft cap on free use and routes heavy users onto Meta’s subscription plans. The pricing model lines up with the same monthly tiers Meta began selling in May, with full creation headroom reserved for paying subscribers after the cap is hit. The cost for the rest of the surfaces is paid by public account holders, whose photos become default source material the moment they leave their profile open.

The Old Partner That’s Stepping Out

Muse Image is also where Meta begins stepping back from a partner. Until Tuesday, image generation inside Meta’s apps ran on third-party models licensed from Midjourney and Black Forest Labs, the two AI image shops Meta struck deals with in 2025 to handle features its own models could not. The change is part of Meta’s shift away from Midjourney as the default image generator inside Meta’s apps. CNBC’s reporting framed Muse Image as the second major release from MSL. The shift trades an outside license fee for an inside compute bill.

Muse Image acts as the creative partner that knows your world, making it easy to turn your ideas into high-quality visuals.

That line came from Meta’s launch blog on Tuesday, the framing Zuckerberg’s AI division wants to leave on the page. The competitive scoreboard does the rest of the work, though. Meta’s internal benchmarks, as reported by CNBC, put Muse Image behind OpenAI’s newest image model, GPT Image 2, in editing tests across single and multiple images. The same benchmarks showed Muse Image ahead of Google’s Nano Banana 2 on those edits, undercutting the consumer breakout Google had with the original Nano Banana last fall.

A $145 Billion Wager Behind the Tool

The price tag behind Muse Image is the largest Meta has ever written. The company expects to spend up to $145 billion on chips, data centers, and other AI infrastructure this year, a range raised from a $115 billion to $135 billion forecast earlier in 2026, according to Meta’s raised capex forecast for 2026. The capex funds MSL, the division Meta renamed a year ago when Wang came in to run it. The roster inside MSL includes an elite research unit called TBD and a recent hardware push that runs alongside the model work. Talent has been the headline cost: the company has offered top AI researchers packages in the hundreds of millions of dollars each.

The money buys more compute than headcount. The capex covers the hardware needed to train and run the next MSL models, and compute is the bottleneck that Wang told staff about at this week’s town hall. Watermelon, the model in line after Muse Spark, uses an order of magnitude more compute than the April release did.

The bet is the response to a year in which Meta fell behind. OpenAI and Alphabet got a head start on consumer image generation, and Google’s Nano Banana became a hit with consumers the previous fall. CNBC’s reporting, citing Meta’s own benchmark testing, put Muse Image behind OpenAI’s GPT Image 2 but ahead of Google’s Nano Banana 2 in the editing tests MSL ran. The benchmark gap is what MSL was built to close.

Muse Spark, the chatbot model MSL shipped in April, performed well on benchmarks but did not match or exceed OpenAI or Anthropic, according to Business Insider. That gap is the one the rest of the capex is meant to fund. Muse Image is one visible delivery against it, and Watermelon is the next one MSL has promised.

The Price Already on the Tab

The same week as the launch, the bill has come due. A bug in a Meta customer-service chatbot, rolled out across Facebook and Instagram in March, let hackers ask the AI agent to swap the email address tied to an Instagram account and trigger a password reset, an exploit captured in the AI-support vulnerability that hit high-profile Instagram accounts. 34,000 Instagram accounts were made vulnerable in the attack, per internal Meta documents viewed by The New York Times. 20,000 of those accounts were breached, exposing email addresses, phone numbers, birth dates, and other personal data.

More than 3,500 of the breached accounts had their usernames taken over and rewritten, the same internal documents showed, and the high-profile casualties included the dormant Obama White House account and a senior official in the US Space Force. Andy Stone, a Meta spokesperson, told The Times that “some of our internal back-end checks failed in this instance, but it wasn’t due to the AI agent itself, and we’ve addressed the underlying cause.” Meta also told Maine’s attorney general in a separate letter that it was conducting a “comprehensive review” of related security issues. The affected accounts have since been secured, the company said. Meta declined to say what data the attackers had viewed or stolen.

On the recovery side, Meta told The Times that the share of hacked accounts successfully recovered by users in the US and Canada rose 30% last year, an improvement the company attributes to the automated support agents it has put in front of users. On the people side, the same Times reporting found Meta had rolled out a program to track employees’ computer activity for AI training, even as it laid off thousands of workers to offset AI spending. Meta did not change its broader AI customer-service plans after the incident, leaving every entrypoint except the IG Forgot Password Chat active.

Two Chips Still in the Bag

Muse Image was not the only thing Meta showed on Tuesday. The company previewed Muse Video, a sibling product built on the same pretraining base as Muse Image and shipping with native audio support, per Meta’s technical breakdown of Muse Video on its AI blog. The blog post bills Muse Video as the second media-generation model developed inside MSL, with the rollout planned for “soon to creators and Meta AI.” The Verge and TechCrunch both logged that Muse Video is “already in development.”

The bigger lift is the next chatbot model. Watermelon, the model after Avocado (the internal name for Muse Spark), is currently in training and uses an order of magnitude more compute than its predecessor, Wang told staff at an internal town hall this week, per Business Insider’s reporting. Wang told the room Watermelon has reached parity with OpenAI’s GPT-5.5 on key benchmarks, an assessment Business Insider called “the clearest sign yet” that Meta’s rebuilt AI unit is closing the gap with OpenAI; Meta declined to comment to the outlet, and OpenAI did not respond to a request for comment.