Meta has introduced the multimodal model Muse Spark 1.1, which is comparable to Opus 4.8 and GPT-5.5 on benchmarks. At the same time, the corporation has opened paid access to its API for external developers — this is the first time Meta is selling access to its own models rather than distributing them for free with open weights.
Agentic Foundation
Muse Spark 1.1 is designed for complex multi-step tasks requiring coordination of multiple applications. The model can act as a main agent, gathering context and distributing subtasks among parallel sub-agents, as well as an executor, returning control when necessary. A context window of 1 million tokens allows it to retain information over long sessions, compressing data and preserving key steps.
Computer Control and Multimodality
The model is trained to work with a desktop in multi-application scenarios with changing conditions. Instead of step-by-step clicking, Muse Spark 1.1 chooses its own strategy: it writes scripts for routine operations and works directly through the interface for simple ones. It also supports multimodality — text, images, and video. Example: the model records a product on video with a smartphone, extracts photos, generates a listing description, and publishes it through a browser on behalf of the user.
Progress in Coding
Developers have reported significant improvements in working with large codebases. Muse Spark 1.1 can diagnose complex bugs, implement features in enterprise systems, and conduct large-scale code migrations. It supports popular agentic frameworks: planning mode, delegation to sub-agents, goal conditioning, and context compression. Early partners include Replit, Cline, and Box.
Benchmarks and Positioning
The model leads in agentic tests: on MCP Atlas — 88.1 points (competitors around 80), on JobBench — 54.7 versus 48.4 for Opus 4.8 and 38.3 for GPT-5.5. However, it lags in coding: on Terminal-Bench 2.0 — 59.0 versus 82.7 for GPT-5.5, 68.5 for Gemini, and 65.4 for Opus 4.8. Meta acknowledges the gap and states it will continue investing in this area.
Pricing and Availability
Access costs: $1.25 per 1 million input tokens and $4.25 per 1 million output tokens. This is lower than Claude Sonnet 4.6 but higher than initial models from OpenAI and Anthropic. Upon registration, developers receive $20 in free credits. The API is compatible with OpenAI SDK and Anthropic Messages formats. The preview is currently available only to developers in the US.
My opinion: Launching a paid API is a strategic shift for Meta, which previously relied on open Llama models. If Muse Spark 1.1 can truly compete with Opus and GPT in agentic scenarios, it could change the balance of power in the AI agent market, especially in the enterprise automation segment. However, the lag in coding is a serious challenge, and without overcoming it, the model risks remaining a niche tool.