How Meta Rebuilt Its AI Lab to Challenge OpenAI and Google
Meta went from the Llama 4 benchmark scandal to six frontier model releases in six months. Here's how Alexander Wang's lab pulled it off.
What happened to Meta’s AI program in the last 18 months?
Meta went from a public benchmarking scandal with Llama 4 in April 2025 to shipping six major model releases in six months, closing out with Muse Wonder and Muse Wonder Full in October. The turnaround followed a roughly $14 billion investment in Scale AI, the hiring of Scale’s founder Alexander Wang to run a new division called Meta Super Intelligence Labs, and tens of billions of dollars a year in data center spending. Independent benchmarks now put Meta’s best model close to Claude Opus 5.5 and ahead of OpenAI’s latest releases on at least one major index.
TL;DR
- Meta’s Llama 4 launch in April 2025 was undercut by a benchmarking controversy after it submitted a specially tuned version of Llama 4 Maverick to a public leaderboard, which didn’t match the model developers could actually download.
- In response, Mark Zuckerberg restructured Meta’s entire AI effort, investing around $14 billion into Scale AI and bringing in founder Alexander Wang to lead a new unit called Meta Super Intelligence Labs.
- Meta shipped six model releases in six months, from Muse Spark in April 2025 through Muse Wonder and Muse Wonder Full in October, alongside a coding agent (Musecode) and an open weight model (Muse Glimmer).
- On the Artificial Analysis intelligence index, Meta’s models climbed from roughly 31 points to 56 points over that stretch, putting Muse Wonder Full just behind Claude Opus 5.5.
- Zuckerberg has said Meta’s infrastructure spending runs $50 to $60 billion, with a single gigawatt-scale cluster costing an estimated $15 billion a year to operate.
- Alexander Wang built the core Muse research team to under 200 people, betting on talent density over headcount, though Zuckerberg has acknowledged he wasn’t sure the bet would pay off.
- Muse Wonder targets personal and professional agents rather than chat or benchmark leaderboards specifically, built around features Meta calls active memory and active focus.
One coffee. One working app.
You bring the idea. Remy manages the project.
What went wrong with Llama 4?
Llama 4 launched in April 2025 as Meta’s attempt to catch up with OpenAI and Google at the frontier. The models themselves landed short of the top labs, which was disappointing enough on its own. But the bigger problem was credibility. Meta had submitted a special, human-preference-tuned version of Llama 4 Maverick to the LMArena leaderboard, a widely watched community benchmark. That version wasn’t the same model available for download. LMArena publicly stated that Meta’s interpretation of its submission policy didn’t match what it expected from model providers.
That distinction matters because leaderboards like LMArena exist specifically to give developers an independent read on model quality. Submitting a tuned variant that doesn’t match the shipped product undermines the entire point of the exercise. The episode became a reference point for skepticism that would resurface later, once Meta started posting benchmark charts of its own.
How did Meta respond, and why was it such a gamble?
Zuckerberg’s response was to tear down the existing structure and start over. In June 2025, Meta invested roughly $14 billion into Scale AI and brought on its founder, Alexander Wang, to head a new division: Meta Super Intelligence Labs. Meta then ran one of the most aggressive recruiting pushes the industry has seen, offering large compensation packages to pull researchers from OpenAI, Google, and Apple. By August 2025, Meta had restructured its whole AI organization around the new lab.
At the time, a lot of the commentary was skeptical, framing it as Meta trying to buy its way to relevance. Zuckerberg has since been candid about how uncertain the bet felt internally. He described being on the verge of committing somewhere in the range of $100 billion to $200 billion and directly asking Wang whether it would work. Wang’s answer, as Zuckerberg told it, was that as of late last year there was no clear sign it would. Meta’s CFO reportedly warned that failure here would be very bad for the company.
The financial scale underpinning all of this is large by any standard. Zuckerberg has put the cost of running a single gigawatt-class data center cluster at around $15 billion a year, with Meta’s total AI infrastructure spending closer to $50 or $60 billion. Meta’s Prometheus cluster in Ohio already exceeds a gigawatt of capacity, and its Hyperion cluster in Louisiana is being scaled toward 5 gigawatts.
How did Meta go from Llama 4 to a frontier model in a year?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Once the new lab was in place, the release cadence became the story. Muse Spark launched in April 2025 as the first model out of Meta Super Intelligence Labs. Muse Spark 1.1 followed in July with a million-token context window and a developer API. August brought Muse Spark 1.2 along with Musecode, Meta’s terminal-based coding agent, and shortly after that an open-weight model called Muse Glimmer, a 30 billion parameter model released under the Apache 2.0 license that can run on a single consumer GPU with 24GB of memory. Muse Spark 1.3 shipped in September. Then in October, Meta released Muse Wonder and Muse Wonder Full, calling them its first true frontier-class models.
That’s six releases in six months, plus a dedicated coding agent and an image model (Muse Image). Tracked against the Artificial Analysis intelligence index, the climb is steep: the original Muse Spark scored around 31, Spark 1.1 hit 34, Spark 1.2 jumped to 40, Spark 1.3 reached 48, and Muse Wonder and Muse Wonder Full landed at 52 and 56 respectively. Artificial Analysis noted that Wonder arrived just six weeks after Spark 1.3, meaning Meta gained roughly 25 points on that index in about six months.
Notably, the team responsible is small. Zuckerberg has described the core research group as around 100 people with infrastructure support around them, and Wang has said Muse was built with a team under 200. The stated philosophy is that talent density matters more than headcount, a contrast with the sprawling teams typical of frontier labs.
How does Muse Wonder actually perform?
On Meta’s own internal benchmark, the personal super intelligence index, Muse Wonder Full scored 59.1%, ahead of Claude Opus 5.5 (55.2%), GPT-6 Astra (54.5%), and Claude Fable 5.1 (49.9%). On Automation Bench, which measures end-to-end business workflows, Wonder Full scored 45.4%, topping every other model Meta tested. On Harvey’s legal agent benchmark it scored 13.1% against Opus 5.5’s 4.2%. On long-horizon coding tests, Deep SWE 1.1 put Wonder Full at 79% and standard Wonder at 78.8%, both ahead of Opus 5.5 (74.2%) and GPT-6 Astra (74.1%). On SWE Marathon, described as an ultra-long-horizon engineering test, Wonder Full scored 57.5% versus 46.3% for Opus 5.5.
Independent benchmarks tell a similar, if more modest, story. Artificial Analysis placed Muse Wonder Full at 56 on its intelligence index, just behind Claude Opus 5.5 at 58, with standard Muse Wonder at 52, roughly level with GPT-6 Astra and GPT 6.1 Soul. That index combines ten separate evaluations, including real-world economic tasks, agentic coding in a terminal environment, scientific programming, expert-level reasoning questions, and a hallucination check, so a high composite score reflects breadth rather than a single strong test.
On the Val index, which emphasizes real professional work, Muse Wonder Full ranked fifth overall at 64.57% accuracy, ahead of every OpenAI model tested, including GPT-6 Astra (63.13%) and GPT 6.1 Soul (61.15%). Val’s own writeup framed this as Meta overtaking OpenAI on that measure. Meta’s stock reportedly rose following the announcement.
Is Muse Wonder actually benchmaxed?
Given what happened with Llama 4, skepticism was inevitable, and it arrived almost immediately. Within hours of launch, users posted demos where Muse Wonder failed on relatively simple generation tasks, prompting accusations that the model had been “benchmaxed,” trained heavily toward specific test formats rather than general capability. The term carries weight here specifically because Meta already has a credibility problem on this exact issue from the Llama 4 leaderboard episode.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
It’s worth separating two different claims. Meta’s internal charts (the personal super intelligence index, Automation Bench, Harvey’s legal benchmark) are Meta’s own composites, and there’s no independent way to verify how representative they are. The Artificial Analysis and Val results are run by third parties using fixed methodologies, which makes them harder to game, though Val listed Muse Wonder Full as having no public API at the time of scoring, limiting how much outside testing has actually happened. Whether the gap between strong leaderboard numbers and underwhelming public demos reflects genuine benchmark overfitting, a narrow focus on agentic and coding tasks over general creative generation, or simply early rough edges in a brand-new model is still being sorted out by the developer community testing it.
What does this mean for Meta’s AI strategy going forward?
The clearest shift is philosophical. Meta built its AI reputation on open-weight Llama models, positioning itself as the open alternative to OpenAI and Anthropic. Muse Wonder breaks from that. Muse Wonder and Muse Wonder Full are closed models, and a previously promised open-weight release of Muse Spark 1.2 still hadn’t shipped as of this launch. Meta is increasingly behaving like the closed labs it used to position itself against, at least for its top-tier models, while still releasing smaller open-weight models like Muse Glimmer.
Zuckerberg’s underlying argument is about control. He’s framed the enormous spending as necessary because only a small number of companies will end up building frontier-level AI, and any company that doesn’t own its own frontier model end to end will be stuck building products on top of someone else’s technology. That logic extends across Meta’s product line, from Instagram and WhatsApp to its smart glasses hardware. Shipping new models every few weeks, rather than once a year, is partly a technical story and partly a defensive one: every month spent behind the frontier is a month competitors get to lock in developer and user habits.
Frequently Asked Questions
What is Meta Super Intelligence Labs?
It’s the AI research division Meta created in mid-2025 after restructuring its AI efforts, led by Alexander Wang following Meta’s investment in Scale AI. It’s responsible for the Muse model family, Musecode, and Meta’s push toward frontier-level AI.
What caused the Llama 4 benchmark controversy?
Meta submitted a specially tuned version of Llama 4 Maverick to the LMArena leaderboard that didn’t match the model publicly available for download. LMArena said this didn’t align with its expectations for how providers should submit models, which damaged trust in Meta’s benchmark claims.
How much has Meta spent on AI infrastructure?
Mark Zuckerberg has stated that Meta’s AI infrastructure spending is in the range of $50 to $60 billion, with a single gigawatt-scale data center cluster costing an estimated $15 billion a year to run.
Is Muse Wonder open source?
No. Unlike earlier Llama models, Muse Wonder and Muse Wonder Full are closed models. Meta has released some open-weight models separately, including the 30 billion parameter Muse Glimmer under an Apache 2.0 license.
How does Muse Wonder compare to GPT and Claude models?
On Meta’s internal benchmarks, Muse Wonder Full leads competing models from OpenAI and Anthropic on several agentic and coding tasks. On the independent Artificial Analysis index it scored just behind Claude Opus 5.5, and on the Val index it outranked all tested OpenAI models, though public demos after launch raised questions about how those scores translate to everyday use.
