Meta just dropped a new AI model, and Mark Zuckerberg is shouting the news from some very un-Zuck-like rooftops.
The Meta founder and CEO is known for many things, but a cozy relationship with traditional journalism is not one of them. Following the reelection of Donald Trump early last year and sporting his newly cultivated Cool Guy look, Zuckerberg said in a video published online that Meta was removing its journalistic-style fact-checkers, replacing them with a user-based verification system similar to X’s Community Notes, which he said was part of Meta’s effort towards “restoring free expression on [its] platforms.” He’s also tended to liken social media to a more open and democratically friendly public forum than the “legacy media,” which, in his view, is fraught with biased censorship. And these days, if he’s going to do a sit-down interview, he tends to call on influencers like Theo Von.
But he’s apparently deemed Meta’s latest model to be a monumental enough event to warrant an interview with Bloomberg. He’s also made an exceedingly rare appearance on X to mark the occasion.
The new model, called Muse Spark 1.1, specializes in agentic tasks and digital tool use, according to Meta. Testing results show that the model outperformed Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5 on four agentic benchmarks, including Humanity’s Last Exam. It lagged behind in coding and multimodal capabilities, though much less so in almost every case than the original Muse Spark, which Meta launched in April.
Models’ “agency”—i.e., their ability to execute long-running tasks with little to no human oversight—has become the major selling point for big tech developers over the past year. Much of those companies’ PR efforts have focused on convincing developers and business customers that in the not-too-distant future, autonomous AI agents will be able to handle the work of individual employees, and maybe not so long after that, of entire teams, perhaps even at some point of entire organizations. (For the time being, though, there are still plenty of bugs to work out; agents are prone to misinterpret instructions and behave unpredictably, sometimes with disastrous results.)
But businesses aren’t going to ditch their Anthropic and OpenAI subscriptions en masse just because Meta’s new model shows promise in a handful of benchmarks. Zuck must be aware of this, which is why he’s playing up the cost factor.
Developers will be able to use Muse Spark 1.1 via a new Meta Model API for free until they reach a certain token limit, after which they’ll be billed on a usage-based system that charges roughly one-quarter of the cost of industry-leading models, Zuckerberg told Bloomberg.
It gels with Meta’s ongoing effort to brand itself as a kind of people’s champ in the age of AI, the company that will democratize access to the technology at a time when its competitors—according to Zuckerberg’s line of thinking—are attempting to control it exclusively for their own financial benefit. “Meta’s vision is to bring personal superintelligence to everyone,” he wrote in a blog post last year. “We believe in putting this power in people’s hands to direct it towards what they value in their own lives. This is distinct from others in the industry who believe superintelligence should be directed centrally towards automating all valuable work, and then humanity will live on a dole of its output.”
Zuckerberg is specifically throwing shade on Anthropic, which he told Bloomberg “is sort of keeping a model for themselves and releasing a kind of simpler version of a model,” referring to that company’s recent rerelease of Fable 5, which is much more heavily guardrailled than the much-hyped Mythos. The claim directly mirrors some fearful predictions that have been circulating online about Anthropic and the prophecy of a “permanent underclass”—a segment of society that can’t afford to use the biggest and best AI models, and as a result has to essentially live under the thumb of those who can. And he might find a receptive audience among some business leaders who have recently said they’re ditching Anthropic due in part to the company’s high token costs.
To call Meta a champion of democratic ideals, though, would be like calling Texaco a pillar of environmental stewardship. The company—previously known as Facebook—has a long, ignominious, and well-documented history of deploying algorithms trained to optimize user engagement above all else. The Facebook algorithm has been shown to have played a role in fueling the ethnic cleansing of Rohingya Muslims in Myanmar in 2017, for example, and Meta is now facing $1.4 trillion in state lawsuits alleging its platforms hooked young users and harmed their mental health.
It’s going to take a lot more than a cheap AI model and a trendy wardrobe for Meta’s CEO to sell himself as the people’s champ in the age of AI, in other words.
READ MORE:
Betting on People’s Worst Instincts Has Kind of Always Been Mark Zuckerberg’s Thing
Read the full article here
