Meta has unveiled its new flagship multimodal model, Muse Spark 1.1. This is not just another update—it marks a strategic shift by the company toward closed, proprietary solutions. Muse Spark 1.1 is positioned as a direct competitor to Opus 4.8 and GPT-5.5, and judging by the benchmarks, it has every reason to be. At the same time, a public preview of the Meta Model API has been launched—the first paid access to its own models for third-party developers.
Agentic Capabilities: Planning and Multitasking
The key feature of Muse Spark 1.1 is its architecture, designed for agentic tasks. The model can act as a main agent that gathers context, creates a plan, and distributes execution among parallel sub-agents. It can work with new tools, MCP servers, and custom skills without prior training. A context window of 1 million tokens allows the model to remember actions, retrieve information from earlier stages of work, and compress context while preserving important steps. This makes it an ideal tool for complex, multi-stage operations.
Computer Use: Adaptation and Strategy
Muse Spark 1.1 is trained to work with a desktop in multi-application scenarios. Instead of step-by-step clicking, it chooses its own strategy: for routine operations, it writes scripts; for simple ones, it works directly through the interface. The model maintains context in long sessions and adapts to unfamiliar interfaces with minimal human involvement. At each step, it can generate batches of actions, significantly speeding up task execution.
Programming: Progress, but Not Leadership
Developers have reported significant progress in the model's work with large codebases. Muse Spark 1.1 can diagnose complex bugs, implement features in enterprise systems, and conduct large-scale code migrations. Popular agentic coding frameworks such as Replit, Cline, and Box are supported. However, in the Terminal-Bench 2.0 benchmark, the model scored 59.0 points, trailing behind GPT-5.5 (82.7), Gemini (68.5), and Claude Opus 4.8 (65.4). Meta acknowledged this gap but stated it will continue investing in this area.
Multimodality and Use Cases
The model works with text, images, and video. Announced capabilities include generating code from visual layouts and detailed video descriptions. Meta demonstrated an example with Facebook Marketplace: the model records a product on video with a smartphone, extracts photos, generates a description, and publishes a listing through the browser on behalf of the user. This is a clear example of how agentic AI can automate complex chains of actions.
Benchmarks and Pricing
In agentic tests, Muse Spark 1.1 leads: on MCP Atlas, it scores 88.1 points versus ~80 for competitors; on JobBench, 54.7 versus 48.4 and 38.3. Access pricing: $1.25 per 1 million input tokens and $4.25 per 1 million output tokens. The API is compatible with OpenAI and Anthropic formats. The public preview is currently available only to developers in the United States.
My expert opinion: Muse Spark 1.1 is not just a model, but a statement of intent. Meta is transitioning from an open-source strategy to building a closed ecosystem, where the API will become the primary source of monetization. If the company can close the gap in coding, it will become a serious competitor to OpenAI and Anthropic. Current results show they have every chance.