- Published on
Meta Releases Most Powerful AI Model Muse Spark 1.3, Coding Ability Surpasses GPT-5.6 Sol
- Authors
- Name
- DeepHorus
Meta Platforms officially released its most powerful artificial intelligence model to date, Muse Spark 1.3, on September 2, 2026. This model is designed to extend agent workflows and enhance coding capabilities, marking a significant step for Meta in the generative AI race. The release follows an urgent industry demand for more efficient and reliable AI programming assistants.
In the release statement, Meta's Chief AI Officer Alexandr Wang stated that Muse Spark 1.3 surpasses OpenAI's previously released GPT-5.6 Sol model in coding capabilities and performs on par with Anthropic's Claude Fable 5.1. This comparison directly establishes Muse Spark 1.3's competitive position among current top-tier AI models. Wang emphasized that the new model represents a major breakthrough for Meta in understanding and executing complex software tasks.
In terms of pricing, Muse Spark 1.3 maintains the exact same strategy as its predecessor, Muse Spark 1.2, aiming to lower the barriers for developer migration and adoption. Through the Meta Model API, developers need to pay $1.25 (approximately 8.4 RMB) per 1 million input tokens, while cached inputs cost only $0.15 (approximately 1 RMB), with output costs set at $4.25 (approximately 28.6 RMB) per 1 million tokens. This stable pricing model helps attract and retain the developer ecosystem.
According to Wang, the adoption of Meta's API platform market is strong, with some developers reaching a weekly usage of trillions of tokens. The new model will gradually be integrated into Meta's core products, including Instagram, Facebook, and the Meta AI Assistant. This signals that Muse Spark 1.3's powerful capabilities will soon reach billions of end users, potentially profoundly changing experiences in social media and everyday digital interactions.
According to Meta's official technical blog post, compared to its predecessor, Muse Spark 1.3 has achieved significant improvements in code generation efficiency. It reduces unnecessary code iterations, generates more concise code, and maintains a clearer overall coding style. Specific data indicates that the model's tool call count has decreased by about 20%, while the number of tokens required to complete the same task has been reduced by 25%. These optimizations mean lower latency and more economical costs.
The core improvement of Muse Spark 1.3 lies in its deep optimization for long-duration agent tasks. The model can collaboratively handle multiple complex workflows in a single long thread and possesses the ability to proactively correct plan gaps and track learning outcomes. When faced with ambiguous instructions, it proactively asks clarifying questions; it requests user assistance when encountering obstacles; and it seeks confirmation before executing critical operations, greatly enhancing the safety and reliability of task execution.
Additionally, the new model has improved reliability in following complex long instructions, better preserving detailed requirements without losing constraints in multi-step tasks. Its multi-task handling capability has also been enhanced, allowing more accurate mapping of new instructions to the correct task context. Most importantly, the model has a stronger awareness of its own capabilities and limitations, which effectively reduces the generation of "hallucination" outputs, making it more trustworthy in practical applications.
Industry analysts point out that the release of Muse Spark 1.3 intensifies competition among AI giants in the "agent AI" field. With its open-source tradition and vast product ecosystem, Meta is attempting to translate advanced AI coding and reasoning capabilities into tangible developer advantages and user experience improvements. The model's actual performance and widespread adoption will be closely monitored by the market in the coming months.
