Meta has made a bold move by introducing its new multimodal model Muse Spark 1.1. This is not just an update — it is a paradigm shift. The company is opening paid access to its proprietary models for the first time through the Meta Model API, signaling serious ambitions in the enterprise AI market.
Architecture for Agents
Muse Spark 1.1 is designed not merely as a chatbot, but as a full-fledged agent system. It can take on the role of a primary coordinator: gathering context, creating a plan, and distributing tasks among sub-agents. A key feature is its ability to work with new tools and MCP servers without the need for prior training. A context window of 1 million tokens allows the model to maintain reasoning threads in long sessions, compressing content and extracting relevant data from history.
Computer Control and Coding
In desktop interaction scenarios, Muse Spark 1.1 demonstrates strategic thinking. Instead of step-by-step clicks, it chooses its own approach: writing scripts for routine tasks or directly interacting with the interface. In programming, the model claims progress in working with large codebases, including diagnosing complex bugs and code migration. However, as benchmarks show, it still lags behind leaders in pure coding. On Terminal-Bench 2.0, it scores 59.0 compared to 82.7 for GPT-5.5.
Benchmarks and Positioning
The results are mixed. In agent tests such as MCP Atlas, Muse Spark 1.1 confidently leads with 88.1 points, surpassing Opus 4.8 and GPT-5.5. This confirms that the model is optimized specifically for multi-agent scenarios. In multimodal tasks, it works with text, images, and video, demonstrating practical use cases like automating product listings on Facebook Marketplace.
Strategic Shift and Pricing
The launch of a paid API marks a departure from the open Llama model strategy. Muse Spark 1.1 is a proprietary model with closed weights. Access pricing is set at $1.25 per 1 million input tokens and $4.25 per output tokens. This is cheaper than Claude Sonnet 4.6 but more expensive than basic OpenAI offerings. The API is compatible with OpenAI and Anthropic formats, simplifying migration for developers. Currently, the service is only available in the US.
Analyst's opinion: Meta is betting on a future where AI agents become the primary interface for interacting with the digital world. While their model currently lags in coding, leadership in agent tasks and aggressive pricing could quickly shift the balance of power in the market. The next model, codenamed Watermelon, could be a true turning point.