Meta Superintelligence Labs on July 9 released Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, alongside a public preview of the Meta Model API — the company’s first paid developer access to its own AI models.
What’s new
Muse Spark 1.1 carries a 1-million-token context window with active context management, and adds multimodal support for images, video and PDFs, according to Meta. The model is designed to act as either a primary agent or a subagent, delegating parts of a task to parallel subagents to cut latency, and Meta says it generalizes to new tools and MCP servers without additional training. On coding, the company points to gains on real-world tasks such as diagnosing bugs, adding features to existing codebases, and large-scale code migrations. The model can also drive computer-use workflows, mixing scripted automation with direct interface control. The Meta Model API that hosts it is built to be OpenAI-compatible, easing adoption for developers already building on rival platforms.
Early access partners include Replit, Cline and Box. Replit chief executive Amjad Masad called it a “complete agentic foundation,” while Box VP of AI products Yashodha Bhavnani said it delivers “enterprise capabilities competitive with today’s leading frontier models.” Meta charges $1.25 per million input tokens and $4.25 per million output tokens, according to Reuters — pricing that lands close to, though slightly above, Anthropic’s Claude Haiku 4.5 and OpenAI’s GPT-5.6 Luna.
Entering a crowded field
The release makes Meta a commercial API vendor for the first time, after years of giving away its Llama models for free. It also lands just a day after OpenAI’s GPT-5.6 family went generally available and the same week SpaceXAI’s Grok 4.5 launched — intensifying a pricing and capability race across frontier labs. Meta has trailed Anthropic and OpenAI specifically in coding-focused models, and chief executive Mark Zuckerberg told employees this month that the company’s broader push into AI agents had been slower than expected.
Safety testing
Meta said Muse Spark 1.1 was evaluated under its Advanced AI Scaling Framework against chemical and biological, cybersecurity, and loss-of-control risk thresholds, and that it performed within safe margins, with improved resistance to jailbreaks and prompt injection and lower rates of hallucination and sycophancy compared with earlier versions.