Meta's last AI model, Llama 4, launched in April 2025 and almost immediately fell apart. The company had submitted a specially tuned, unreleased version to a public AI scoring contest, which ranked it near the top. When the actual publicly available model was tested independently, it dropped to around 32nd place on the same leaderboard. Meta's own outgoing chief AI scientist, Yann LeCun, confirmed in a Financial Times interview that the results had been manipulated, saying the team "used different models for different benchmarks to give better results." Zuckerberg reportedly lost confidence in everyone involved and effectively dismantled the entire AI division responsible.
The rebuild was expensive and fast. Meta paid $14.3 billion for a 49% stake in Scale AI, a company that specialises in preparing data to train AI systems, and brought its founder, Alexandr Wang, to Meta as its first-ever chief AI officer. Wang was 28 at the time. LeCun, the man he replaced in terms of influence, publicly called him "young" and "inexperienced," though he acknowledged Wang "learns fast" and "knows what he doesn't know." Wang then hired heavily from OpenAI, Anthropic, and Google, reportedly offering signing packages worth hundreds of millions of dollars.
Nine months later, Muse Spark is the output. Independent evaluators place it fourth globally, scoring 52 on the most comprehensive third-party AI ranking, behind Google Gemini at 57, OpenAI's ChatGPT at 57, and Claude at 53. That is a real gap from the leaders, but not a disqualifying one. Where Muse Spark genuinely leads the field is health reasoning: working with over 1,000 physicians to build training data, Meta produced a model that outperforms every paid competitor on medical and health-related questions. For professionals in healthcare, insurance, or any sector that touches health data, that is the most notable specific capability in this launch.
The weaker areas are real and worth knowing. Muse Spark trails significantly on abstract reasoning tasks and on the kind of complex, multi-step work where you hand the AI a job and walk away. On autonomous task handling, it sits well behind ChatGPT and Claude. If your team is evaluating AI for back-office automation or any workflow where the AI needs to complete a sequence of tasks independently, the leading paid models still hold a clear advantage.
The strategic move underneath all of this is the price. Muse Spark is completely free. ChatGPT and Claude charge $20 per month for full access; the most powerful ChatGPT mode costs $200 per month. Meta can offer its model for free because its revenue comes from advertising, not subscriptions. The catch is that to use Muse Spark you need a Facebook or Instagram account, which means your queries are connected to the same profile Meta uses for ad targeting. That is a genuine privacy consideration for anyone using AI to work through sensitive business questions.
The other notable shift is that, unlike Meta's previous AI models which were freely downloadable and modifiable by anyone, Muse Spark is closed and proprietary. Meta says it hopes to open-source future versions but has made no commitment. This matters for businesses that had been building products on Meta's open models: that era may be ending.
Muse Spark is currently rolling out across the Meta AI app, WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban smart glasses. For most businesses, the practical implication is simpler than the technical story: a free, capable AI assistant is about to appear inside the apps your customers and your staff already use every day. Whether that helps or complicates your own AI plans depends on what you are trying to do with AI. But it does mean the baseline expectation for what AI should cost, and where it should be available, just shifted.