Meta introduces Muse Spark 1.1: betting on agents, coding, and a paid API
Meta has taken a significant step in its AI strategy by releasing the multimodal model Muse Spark 1.1. According to stated benchmarks, the new model demonstrates results on par with Opus 4.8 and GPT-5.5. However, the key event is the launch of the public preview of the Meta Model API. This is the first time Meta has provided paid access to its proprietary models for external developers, signaling a shift in the company's course from open Llama models to the commercialization of closed solutions.
Agent Architecture and Context
Muse Spark 1.1 is designed as a full-fledged agent system. It can plan and coordinate the work of multiple applications and services. The model works with new tools, MCP servers, and custom skills without the need for prior training. In the main agent mode, it gathers context, builds a plan, and distributes tasks among parallel sub-agents. The context window is 1 million tokens, allowing the model to remember actions from early stages of work and effectively compress context.
Computer Control and Coding
The model is trained to work with the desktop in multi-application scenarios, adapting to unfamiliar interfaces. Instead of step-by-step clicks, it chooses its own strategy: for routine operations, it writes scripts, and for simple ones, it works directly through the interface. In the field of programming, Muse Spark 1.1 shows progress in working with large codebases, including diagnosing complex bugs and code migration. Popular agentic coding frameworks are supported, with early partners including Replit, Cline, and Box.
Multimodality and Safety
The model works with text, images, and video. In a demonstration, Meta showed a Facebook Marketplace scenario: the model records a product on video, extracts photos, generates a description, and publishes the listing. In terms of safety, the model has been evaluated under the Advanced AI Scaling Framework protocol and falls within acceptable limits across all categories of frontier risks.
Benchmarks and API
In agent tests, Muse Spark 1.1 leads: on MCP Atlas — 88.1 points versus ~80 for competitors, on JobBench — 54.7 versus 48.4 and 38.3. However, in coding, the model lags behind: on Terminal-Bench 2.0 — 59.0 versus 82.7 for GPT-5.5. Meta acknowledges the gap and is investing in this area. API pricing: $1.25 per million input tokens and $4.25 per million output tokens. The API is compatible with OpenAI and Anthropic formats, simplifying migration.
Expert Opinion
The launch of the paid API represents a strategic pivot by Meta from open models to proprietary ones. The company is clearly aiming to compete with OpenAI and Anthropic in the agent solutions segment, but the lag in coding remains a vulnerability. Given that Meta is already training a more powerful model under the codename Watermelon, we are witnessing the beginning of a race where agent capabilities will become the main battlefield.