Mark Zuckerberg broke a three year silence on X to announce Muse Spark 1.1 had landed on OpenRouter. The fine print says only developers with a US address can actually use it.
Zuckerberg hadn't posted on X since 2023. On July 9 he came back for one line: "A lot of people asked for this, so Muse Spark 1.1 is now on OpenRouter." No thread, no victory lap, just the announcement. Meta's newest coding model had finally reached the marketplace where developers actually shop for models instead of being boxed inside Meta's own apps. That part was good news. Then came the catch nobody asked for: access is a public preview, and it's US developers only, according to Meta's own developer announcement.
Muse Spark 1.1 is the second release out of Meta Superintelligence Labs, the unit Alexandr Wang has run since Meta poached him out of Scale AI. The first Muse Spark model launched in April, according to Fortune. This one is open weight and multimodal, built for agentic and coding tasks, and priced to undercut the competition hard. Meta charges $1.25 per million input tokens and $4.25 per million output tokens, roughly a quarter of what Anthropic and OpenAI charge for comparable models, and new accounts get $20 in free credits before the meter starts. That's an aggressive number. It's also table stakes if you actually want developers to switch.
Meta says Muse Spark 1.1 beats OpenAI and Anthropic models on certain tasks. On SWE-Bench Pro, though, it scored 61.5. Anthropic's Opus 4.8 scored 69.2. OpenAI's GPT-5.5 scored 83.4. Meta's own comparisons skipped the most recent releases from both rivals, which is a convenient way to win a benchmark chart. Then Hacker News found something messier: a commenter alleged Meta ran its Terminal-Bench 2.1 numbers using six CPU cores, on a benchmark that caps tasks at four. "This disqualifies the results," the commenter wrote. Meta hadn't responded to the allegation as of this writing. Wang, for his part, has called Muse Spark an appetizer, which reads less like humility and more like an admission that this isn't the model meant to win the argument.
Frankly, the OpenRouter move was the right call. Meta had been keeping Muse Spark's API bottled up inside its own properties, which meant developers who wanted to test it against Claude or GPT in their existing workflows couldn't, at least not without extra plumbing. OpenRouter exists precisely to remove that friction: one interface, dozens of models, no separate account needed for each one. Putting Muse Spark 1.1 there was Meta finally meeting developers where they already work.
One step forward, a few steps back #
The US-only restriction undoes a chunk of that goodwill immediately. Meta built its AI reputation on Llama being open to the entire world, weights downloadable from Seoul to São Paulo. A paid API that only clears US addresses is a different animal, and it sits awkwardly next to that history. Meta hasn't announced a timeline for international availability. Developers in London, Lagos, or Bangalore who want to build on Muse Spark 1.1 through OpenRouter are simply out of luck for now.
There's a pattern here worth naming. Meta ships something genuinely useful, then attaches a condition that narrows who benefits. Open weights, but the fastest model stays gated. A marketplace listing, but only for one country's developers. Wang's superintelligence unit is still young, barely four months removed from its first model. The ambition is real and the pricing proves Meta wants developers, not just headlines. But the same week Zuckerberg ended a three year silence to announce broader access, Meta drew a line around exactly who gets it. Whether that line moves depends on what Meta says next, and so far, it hasn't said anything at all.
Also read: A Former DeepMind Researcher Just Raised $55 Million Before Building Anything • Anthropic Turns Claude Code's Throwaway Dashboards Into Live Internal Tools • Roblox Lets Anyone Build a Game From a Text Prompt on Their Phone