Meta released Muse Spark 1.1 on 9 July 2026, its second model from Meta Superintelligence Labs and a multimodal reasoning model built for agentic work — tool use, computer use, coding and orchestration across apps. It is available now in a public preview: free in a Thinking mode inside the Meta AI app and at meta.ai, and, for the first time, through a paid Meta Model API aimed at developers. The benchmark scores will get the attention, but they are the least interesting thing here. The significant move is that Meta has put a meter on a closed model.

The pivot is the story

For three years, Meta's AI strategy was defined by giving models away. The open-weight Llama family was the counter-move to a market of paid, closed APIs — Meta's argument that frontier capability should be free to download and run. Muse Spark 1.1 is the clearest signal yet that Meta believes frontier AI economics require a different commercial model. The model is proprietary and closed-weight, it is served through a metered API, and it is the first Meta-developed frontier model the company has put behind a paid API. Meta says Muse-family models are expected over time to replace the Llama models currently powering the assistants inside WhatsApp, Instagram and Facebook. A company that spent three years undercutting paid APIs with free weights has now built one of its own.

That shift sits inside a larger reorganisation. Meta Superintelligence Labs, the unit behind the model, is run by Alexandr Wang, brought in as Meta's first chief AI officer after the company took a $14.3 billion, 49% non-voting stake in his data-labelling firm Scale AI in 2025. Muse Spark 1.1 is the second model out of that reorganised unit — the original Muse Spark arrived in April 2026 — and it landed two days after Muse Image, the labs' first image-generation model. The cadence is the point: Meta is shipping paid, closed products at speed after a public restructuring of its entire AI effort.

What developers actually get

The Meta Model API launched in public preview for US developers, and Meta built it to be compatible with the widely used OpenAI API format, which lowers the switching cost for teams already wired into that ecosystem. Pricing is $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits at sign-up — rates that sit well below the flagship pricing charged by OpenAI and other leading frontier developers, which is the entire pitch. Mark Zuckerberg framed it plainly in a social post, describing the goal as a strong agentic and coding model at a very low price.

On capability, Muse Spark 1.1 is a natively multimodal model accepting text, image, video, PDF and audio input and returning text, running a one-million-token context window with built-in compression — a large jump from the roughly 260,000 tokens of the original. It is also built for multi-agent orchestration natively: it can act as a primary agent delegating to subagents, or as a subagent itself, and its computer-use mode decides on its own whether to write a script or click through an interface. Those agentic capabilities are where Meta's own figures look strongest.

Read the benchmarks with care

Meta's published numbers are selectively strong, and that selection matters. On agentic evaluations the model posts genuinely competitive results — a reported 88.1 on MCP Atlas, 54.7 on JobBench, 62.1 on Humanity's Last Exam with tools. On coding, the picture is weaker by Meta's own tables: 61.5 on SWE-Bench Pro and 80.0 on Terminal-Bench 2.1 trail the leading frontier models, even as they improve substantially on the original Muse Spark. This is a model that is stronger at orchestrating tools than at writing difficult code, and Meta's materials are arranged to lead with the former.

Two caveats belong next to any of these figures. First, they are Meta's own launch benchmarks; independent testing has already scored the model around ten points below Meta's figure on at least one coding benchmark, a reminder that vendor evals are an opening argument, not a verdict. Second, the low headline price comes with a wrinkle: reasoning tokens are billed at the full output rate, so an agentic model that thinks extensively before acting can cost more per task than the per-token rate suggests. The parameter count, architecture and knowledge cutoff are undisclosed. On safety, Meta says the model was evaluated under its Advanced AI Scaling Framework and reports jailbreak resistance and safe margins on chemical, biological, cyber and loss-of-control risks — claims that are the company's own and unverified externally.

Key Takeaways

  • Meta released Muse Spark 1.1 on 9 July 2026 in public preview — its second Superintelligence Labs model and the first Meta-developed frontier model it has put behind a paid API, proprietary and closed-weight.

  • It launched with the new Meta Model API (OpenAI-compatible, US developers) at $1.25/$4.25 per million input/output tokens plus $20 in free credits.

  • The real story is strategic: Meta is adding a metered, proprietary model alongside the open-weight Llama strategy that defined its earlier AI push, and expects Muse-family models to take over its consumer assistants over time.

  • Meta's benchmarks are strongest on agentic tasks and weaker on coding; independent testing already shows a gap on at least one coding benchmark.

  • Reasoning tokens bill at the full output rate, so the low per-token price needs checking against real cost-per-task; parameters, architecture and cutoff are undisclosed.