Meta has introduced the multimodal model Muse Spark 1.1, which benchmarks at the level of Opus 4.8 and GPT-5.5. Simultaneously, the company launched a public preview of the Meta Model API—the first paid access to its proprietary models for external developers.
Agent Capabilities
The model is designed for tasks requiring the planning and coordination of multiple applications and services. Muse Spark 1.1 can work with new tools, MCP servers, and user skills without prior training.
As the main agent, the solution gathers context, creates a plan, and distributes tasks among parallel sub-agents. As a sub-agent, it performs its task and returns control to the main agent if necessary.
The context window is 1 million tokens. The model remembers actions, retrieves information from earlier stages, and compresses context while preserving user-important steps.
Computer Usage
Muse Spark 1.1 is trained to work with desktops in multi-application scenarios with changing conditions. The model maintains context in long sessions and adapts to unfamiliar interfaces with minimal human intervention.
Instead of step-by-step clicking, it chooses a strategy. For routine operations, the neural network writes scripts; for simple ones, it works directly through the interface. At each step, it can generate action packages.
Programming
Developers reported significant progress in working with large codebases. Muse Spark 1.1 can diagnose complex bugs, implement features in enterprise systems, and conduct large-scale code migrations.
The technology also supports popular agent frameworks for coding: planning mode, delegation to sub-agents, goal conditioning, and context compression. Early partners include Replit, Cline, and Box.
Multimodality
The model works with text, images, and video. Its capabilities include generating code from visual layouts, detailed image and video descriptions, and agent scenarios where perception and action occur simultaneously.
Meta demonstrated an example with Facebook Marketplace: the model records a product on video with a smartphone, extracts photos, creates a listing description, and publishes it through a browser on behalf of the user.
Benchmarks
Meta published a comparison table with Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. The results are mixed.
In agent tests, Muse Spark 1.1 leads. On MCP Atlas (extensive tool usage), the model scored 88.1 points—Opus 4.8 and GPT-5.5 scored around 80 points. On JobBench (professional tool usage)—54.7 versus 48.4 for Opus 4.8 and 38.3 for GPT-5.5.
In coding, the model lags behind leaders. On Terminal-Bench 2.0, Muse Spark 1.1 scored 59.0. For comparison: GPT-5.5 scored 82.7, Gemini 68.5, and Claude Opus 4.8 scored 65.4. Meta acknowledged the gap and stated it will continue investing in this area.
On the company’s internal benchmark, the model significantly improved over the first Muse Spark and, according to the company, competes with leading alternatives.
Security
Meta conducted a security assessment using the Advanced AI Scaling Framework. Across all frontier risk categories—chemical and biological threats, cybersecurity, and loss of control—the model is within acceptable limits.
The company reported resilience to direct jailbreaks and attacks through untrusted data, prompt injections, and system prompt attacks. Developers also noted reduced hallucination and parroting levels.
API and Pricing
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. Developers receive $20 in free credits upon registration.
The API is compatible with OpenAI SDK formats (Chat Completions and Responses) and Anthropic Messages. To connect, simply change the base URL to api.meta.ai/v1 and specify the key without rewriting code.
The public preview is currently available only to developers in the U.S. For clients, the model operates in Thinking mode in the Meta AI app and on the meta.ai website.
Strategic Shift
The launch of the paid API marks a strategic shift for Meta. Previously, the company built its AI strategy around open Llama family models with public weights. Muse Spark 1.1 is a proprietary model with closed weights.
Meta’s AI Director Alexander Wong told CNBC that improving coding capabilities was a key focus of the update. According to him, strong coding capabilities are essential for creating full-fledged AI agents. The company is already training a more powerful model codenamed Watermelon. Release dates are not disclosed.
The first Muse Spark was released in April and was available only to select partners in a private preview.
Earlier, on July 7, Meta launched the Muse Image generation model and showcased an early preview of Muse Video.
