Meta has taken a significant step toward commercializing its AI developments by introducing the new multimodal model Muse Spark 1.1. The company not only claims the new model is competitive with Opus 4.8 and GPT-5.5, but it is also opening paid access to its models for the first time through the public preview of the Meta Model API. This marks a strategic shift from an open-source policy to proprietary solutions.
Next-Generation Agent Architecture
Muse Spark 1.1 is designed as a full-fledged agent system. The model is capable of planning, coordinating the work of multiple applications and services, and interacting with MCP servers and user skills without the need for prior training. In main agent mode, it gathers context, forms a plan, and distributes tasks among sub-agents, which return control back when necessary. A context window of 1 million tokens allows the model to effectively work with long sessions and retrieve information from early stages of task execution.
Coding: Lagging Behind, but Ambitious
Meta emphasizes improvements in programming skills, as confirmed by the company's AI director. Muse Spark 1.1 shows progress in working with large codebases, diagnosing complex bugs, and conducting large-scale migrations. However, on benchmarks such as Terminal-Bench 2.0, the model (59.0 points) still lags behind GPT-5.5 (82.7) and Opus 4.8 (65.4). The company acknowledges this gap and is already training a more powerful model under the codename Watermelon. Nevertheless, support for popular coding frameworks and partnerships with Replit, Cline, and Box indicate Meta's serious intentions in this segment.
Multimodality and Safety
The model works with text, images, and video. One illustrative example is automating listings on Facebook Marketplace: the model captures a product on video, extracts photos, generates a description, and publishes it. In the area of safety, Meta claims resilience to jailbreaks and prompt injections, as well as a reduced level of hallucinations. Evaluation under the Advanced AI Scaling Framework protocol showed that risks are within acceptable limits.
Pricing and Availability
The API cost is $1.25 per 1 million input tokens and $4.25 per 1 million output tokens, positioning the model between budget solutions and the premium segment. Developers receive $20 in free credits upon registration. The API is compatible with OpenAI and Anthropic formats, simplifying migration. Access is currently limited to developers in the United States.
My opinion: The launch of Muse Spark 1.1 and the paid API is not just a model update, but a paradigm shift for Meta. The company, which previously bet on open-weight Llama, is now clearly aiming to secure a place in the league of commercial AI giants. However, the lag in coding is a critical shortcoming for agent systems, and the success of Watermelon will be a decisive factor in this race.