Meta is making a strategic pivot, introducing Muse Spark 1.1 to the world — a multimodal model that demonstrates competitiveness with Opus 4.8 and GPT-5.5 on benchmarks. Simultaneously, the company is launching a public preview of the Meta Model API, marking the first step toward monetizing its own AI models for external developers. This is a fundamentally new course for the company, which previously focused on open models from the Llama family.
Agent Architecture and Computer Interaction
Muse Spark 1.1 is designed for complex agent tasks requiring planning and coordination across multiple applications. The model can act as a main agent, gathering context, creating a plan, and distributing tasks among sub-agents, as well as in an executor role. A context window of 1 million tokens allows it to remember actions and retrieve information from earlier stages of work.
Particularly noteworthy is its ability to work with the desktop in multi-application scenarios. Instead of step-by-step clicking, Muse Spark 1.1 chooses its own strategy: it writes scripts for routine operations, works directly through the interface for simple tasks, and adapts to unfamiliar conditions with minimal human involvement.
Programming and Multimodality
The developers claim significant progress in coding. The model can diagnose complex bugs, implement features in enterprise systems, and conduct large-scale code migrations. It supports popular agent frameworks, with early partners including Replit, Cline, and Box. Multimodality allows it to work with text, images, and video, including generating code from visual layouts.
Benchmarks and Positioning
Test results are mixed. In agent tasks, Muse Spark 1.1 confidently leads: on MCP Atlas, it scored 88.1 points versus ~80 for competitors. However, in coding, the model lags behind: on Terminal-Bench 2.0, the result is 59.0 versus 82.7 for GPT-5.5. Meta acknowledges this gap and is already training a more powerful model under the codename Watermelon.
Safety and API
Under the Advanced AI Scaling Framework protocol, the model falls within acceptable limits across all categories of frontier risks. API pricing is set at $1.25 per 1 million input tokens and $4.25 per output token, lower than Claude Sonnet 4.6. The API is compatible with OpenAI and Anthropic formats, simplifying developer migration.
Expert opinion. The launch of Muse Spark 1.1 is not just another model update, but a clear signal of a shift in Meta's strategy. The transition from open Llama to proprietary Muse Spark with a paid API indicates the company's desire to secure a place in the elite segment of AI agents. However, the lag in coding is a critical weakness that needs to be addressed if Meta wants to seriously compete with OpenAI and Anthropic in this area. For now, this is more of a bid for participation than a victory.