Meta has unveiled its 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, marking the first paid access to its models for external developers.

Agent Capabilities

The model is designed for tasks requiring planning and coordination across 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 executes its task and returns control to the main agent when necessary.

The context window spans 1 million tokens. The model remembers actions, extracts information from earlier stages, and compresses context while retaining important steps for the user.

Computer Usage

Muse Spark 1.1 is trained to operate on desktops in multi-application scenarios with changing conditions. The model maintains context during long sessions and adapts to unfamiliar interfaces with minimal human intervention.

Comparison of Muse Spark 1.1's capabilities with other models in computer usage. Source: Meta.

Instead of step-by-step clicking, it autonomously selects strategies. For routine tasks, the neural network writes scripts, while for simpler tasks, it interacts directly through the interface. At each step, it can generate action packets.

Programming

Developers have reported significant progress in the model's ability to handle large codebases. Muse Spark 1.1 can diagnose complex bugs, integrate features into corporate systems, and conduct large-scale code migrations.

Comparison of Muse Spark 1.1 and Muse Spark in programming capabilities. Source: Meta.

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. Among its capabilities are generating code from visual layouts, providing detailed descriptions of images and videos, and executing agent scenarios where perception and action occur simultaneously.

Meta demonstrated an example with Facebook Marketplace: the model captures a product on video with a smartphone, extracts photos, generates a listing description, and publishes it through the browser on behalf of the user.

Benchmarks

Meta released a comparative table featuring Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. The results vary.

Muse Spark 1.1's results across various benchmarks. Source: Meta.

In agent tests, Muse Spark 1.1 leads the pack. On MCP Atlas (extensive tool usage), the model scored 88.1, while Opus 4.8 and GPT-5.5 scored around 80. On JobBench (professional tool usage), it scored 54.7 compared to 48.4 for Opus 4.8 and 38.3 for GPT-5.5.

In coding, the model trails behind the leaders. On Terminal-Bench 2.0, Muse Spark 1.1 scored 59.0, while GPT-5.5 scored 82.7, Gemini 68.5, and Claude Opus 4.8 65.4. Meta acknowledged this shortfall and stated it will continue investing in this area.

On the company's internal benchmark, the model significantly improved upon the first Muse Spark and, according to the company, competes with leading alternatives.

Safety

Meta conducted a safety assessment using the Advanced AI Scaling Framework. Across all categories of frontier risks—chemical and biological threats, cybersecurity, and loss of control—the model remains within acceptable limits.

The company reported resilience against direct jailbreaking and attacks via untrusted data, prompt injections, and attacks on system prompts. Developers noted a reduction in hallucinations and acquiescence.

API and Pricing

Access costs $1.25 per million input tokens and $4.25 per million output tokens. This is lower than Claude Sonnet 4.6 but higher than the entry-level models from OpenAI and Anthropic. Upon registration, developers receive $20 in free credits.

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 provide 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 within the Meta AI app and on the website meta.ai.

Strategic Shift

The launch of the paid API marks a shift for Meta. Previously, the company built its AI strategy around open models from the Llama family with public weights. Muse Spark 1.1 is a proprietary model with closed weights.

Meta's AI director, Alexander Wong, stated to CNBC that improving coding capabilities was a key focus of the update. He emphasized that without strong coding capabilities, it is impossible to create fully functional AI agents. The company is already training a more powerful model codenamed Watermelon, with no release timeline disclosed.

The first Muse Spark was released in April and was available only to select partners in a private preview.

Notably, on July 7, Meta launched the image generation model Muse Image and provided an early preview of Muse Video.