Meta opened a public preview of its Meta Model API on July 9, giving outside developers programmatic access to Muse Spark for the first time. The company paired the launch with Muse Spark 1.1, an upgraded version of the agentic reasoning model it introduced earlier this year. Until now, Muse Spark existed only inside Meta’s own consumer surfaces, with no committed timeline for third-party API access.
That gap mattered. Meta had positioned Muse Spark as a serious agentic and coding model, the kind enterprises evaluate against offerings from OpenAI, Anthropic, and Google. Without an API, that positioning was aspirational. Rivals had already opened equivalent access months earlier. The public preview turns Muse Spark from a showcase feature into a product developers can actually build against, priced and integrated like the rest of the field.
Muse Spark 1.1 itself targets three areas: agent orchestration, computer use, and coding. According to Meta, the model can act as a main agent that plans and delegates to parallel subagents, or as a subagent that executes a bounded task and escalates when it hits a limit. It manages a context window of 1 million tokens, retrieving details from earlier in a session rather than losing them to compaction. Meta says the model decides when to script an action versus click through an interface directly, a distinction that matters for computer-use agents operating across multiple applications where conditions change mid-task.
On coding, Meta reports substantial gains over the original Muse Spark on tasks involving large, established codebases: bug diagnosis, feature work inside enterprise systems, and multi-step migrations. The company did not publish independent benchmark comparisons against competing coding models, relying instead on its internal Meta Internal Coding Bench and testimonials from early partners. Replit CEO Amjad Masad and Cline CEO Saoud Rizwan both praised the model’s tool calling and coding depth in statements included with the release, and Box VP of AI Products Yashodha Bhavnani cited results from the company’s internal enterprise evaluation set. Testimonials from launch partners are not a substitute for third-party leaderboard results, and none accompanied this release.
Meta says it ran the model through its Advanced AI Scaling Framework, evaluating chemical and biological risk, cybersecurity, and loss of control before deployment, and reports Muse Spark 1.1 operates within safe margins on all three with improved resistance to jailbreaks and prompt injection. The full evaluation report is published alongside the release, though the underlying test methodology is Meta’s own.
The timing lines up with Muse Image, a separate model Meta shipped the same week, suggesting a coordinated push to establish Meta Superintelligence Labs as a multimodal platform vendor rather than a single-model lab. An API without pricing details or committed rate limits is still a preview, not a product. Developers evaluating agentic coding stacks should treat Muse Spark 1.1 as a candidate worth a benchmark run against their own workloads, not a settled choice, since the comparative data needed to rank it against Claude, GPT, and Gemini on identical tasks does not yet exist in public form.
Reported by Meta on July 9, 2026.
