Meta has a new flagship AI model, and for the first time, it comes with a price tag. Muse Spark 1.1, announced Thursday, is a multimodal reasoning model built for agentic tasks — and per Bloomberg, the first Meta model to ship with a paid tier for developers.
What the model does
Meta's pitch centers on autonomy. Muse Spark 1.1 posts what the company calls major gains in tool and computer use, coding, and multimodal understanding, and it "advances the performance-efficiency frontier." Translated: it's built to operate software, not just chat.
The details are more interesting than the benchmarks. The model holds context across sessions, adapts when requirements shift mid-task, and can find its way around interfaces it's never seen with minimal hand-holding. Meta says it trained the model to make a judgment call humans make constantly — write a script when automation is faster, click when clicking is simpler, and batch actions together when that's the efficient move.
It landed the same week as Muse Image, the company's new image model, and Meta framed both as steps toward its "personal superintelligence" vision. Vishal Shah, Meta's VP of AI products, added that the company is building tools that act on a user's behalf across outside apps and services. The company claims the model beat many competitors in coding and agentic benchmarks — Meta's own numbers, worth the usual grain of salt.
The first bill Meta's AI has ever sent
The strategic shift is the paid tier. Mark Zuckerberg told Bloomberg the developer pricing will be "very aggressive and attractive" — Meta undercutting rivals to buy its way into a market where OpenAI, Anthropic, and Google already collect revenue.
He needs the pathway more than the headline suggests. Meta has committed hundreds of billions of dollars to AI — by some accounts north of a trillion — while roughly 98% of its ~$200 billion in 2025 revenue still came from ads. Even at an implausible $100 billion a year in AI subscriptions, the math on breaking even runs past a decade. Every paid tier from here is Meta trying to build a second business as large as its first.
The uncomfortable footnote
There's a reason to pause before handing this model your workflows. Just last month, reports surfaced that Meta's agentic support bots on Instagram had incorrectly granted hackers access to roughly 20,000 accounts — after the hackers simply asked. That's the same category of AI Meta is now selling: models empowered to take actions on someone's behalf.
Meta is confident Muse Spark 1.1 behaves better. Maybe it does. But the company is also planning to replace 90% of its content review staff with AI by year's end, so its incentive to believe its own agents is enormous. Whether the model earns that trust — from developers being asked to pay for it — is the actual test that starts now.
An impressive model, a necessary business move, and a trust deficit, all in one launch. That's Meta's AI story in miniature.
Image: BM Amaro, via Pexels
