Meta has been lagging in AI over the last couple of years, and when it’s finally found its footing, CEO Mark Zuckerberg has shown up where all the AI action is.
Zuckerberg posted on X on Thursday from his old @finkd handle, a account he’d more or less abandoned since 2023, to announce Muse Spark 1.1, Meta’s latest general-purpose AI model built for coding and agentic work. “Today we’re releasing Muse Spark 1.1 — a strong agentic and coding model at a very low price. It’s available through our new Meta Model API and in Meta AI,” he wrote, in a post that would’ve been unremarkable coming from almost anyone else, but felt notable given who was doing the posting and where.
To understand why this matters, it helps to go back to what Meta has been through. The company’s last major model release before its current run was Llama 4, and that launch turned into one of the messier moments in Meta’s AI history. Former Meta Chief Scientist Yann LeCun later admitted that the team had fudged the Llama 4 benchmarks “a little bit,” mixing different versions of the model to hit better numbers on different tests. Zuckerberg was reportedly furious once this came out, and effectively sidelined the GenAI organization responsible. Meta acquired Scale AI soon after, installed its CEO Alexandr Wang to lead a rebuilt AI effort, and laid off hundreds of engineers along the way. The company also quietly walked back its open-source commitments, a year after Zuckerberg had written an entire essay on why open-sourcing AI was the right long-term bet.
What’s followed since is Wang’s Meta Superintelligence Labs putting out a steady run of releases meant to close that gap. The lab put out its first model, Muse Spark, in April, and this week it’s followed that up with Muse Image and Muse Video, and now Muse Spark 1.1. Meta says Muse Spark 1.1 is a significant jump over the original model, with the biggest gains coming in tool use, computer use, coding, and how it handles images and video together. It can manage a context window of a million tokens on its own, deciding what to hold onto, what to pull back up, and what to compress as a session runs long, which Meta is positioning as core to how the model handles agentic tasks: acting as a main agent that plans out work and hands pieces of it off to subagents running in parallel.
The benchmark numbers back up a lot of that framing. On MCP Atlas, a test of scaled tool use, Muse Spark 1.1 scores 88.1, ahead of both Claude Opus 4.8 and GPT 5.5. On JobBench, a professional tool-use benchmark, it posts 54.7 against Opus 4.8’s 48.4 and GPT 5.5’s 38.3, which is the widest gap in the entire release. It also leads on Humanity’s Last Exam with tools and on Finance Agent v2. Coding is where the picture is closer to parity than dominance — Muse Spark 1.1 trails GPT 5.5 and Opus 4.8 on Terminal-Bench 2.1, SWE-Bench Pro, and DeepSWE 1.1, though it’s still a huge jump over the original Muse Spark’s numbers on the same tests. Where Meta makes its clearest case is cost. Plotted against price per task, Muse Spark 1.1 comes close to Opus 4.8’s ceiling on OSWorld 2.0 at a fraction of what it costs to run there.
Muse Spark 1.1 arrives two days after Meta rolled out Muse Image and Muse Video, its first media generation models from the same lab. Muse Image landed the No. 2 spot on Arena’s text-to-image leaderboard, behind only OpenAI’s GPT Image 2, and holds that same No. 2 position in single-image and multi-image editing. Muse Video came in at No. 3 on the text-to-video leaderboard, behind Google and ByteDance, with native audio generation built in as a differentiator most rivals don’t yet offer. Between the two releases, Meta has put out a genuinely competitive lineup within the span of a week, something that would’ve seemed unlikely even a few months ago.
None of this happened in isolation either. Meta’s rollout landed on the same day SpaceXAI and Cursor put out Grok 4.5 with its own claims against Opus 4.8 and GPT 5.5, which is a fair snapshot of how compressed the release calendar has gotten across every lab right now.
And that brings the story back to where Zuckerberg chose to post it. Meta has spent years trying to get people to have their public conversations on Threads instead of X, building out quote posts, trending topics, and all the other scaffolding that made X what it is. But when Zuckerberg had actual news to share, the kind of news he wanted developers, researchers and the AI press to see immediately, he posted it on X instead. Threads has audience. X still has the room where AI people are actually arguing, dunking, and paying attention in real time, and that’s the room Zuckerberg needed for this.