Meta released first AI models at flagship ChatGPT and Claude level
American technology company Meta has introduced a new generation of artificial intelligence models, Muse Wonder and Muse Wonder Full, which, according to the developer's estimates, are capable of competing with the most powerful OpenAI and Anthropic systems, in particular GPT-6 Astra and Claude Fable 5.1.
The new products are designed for programming, creating applications, and performing complex multi-step tasks without constant user intervention.
The company announced the launch on October 12 on its official website. These are the most powerful AI models that Meta has presented after a large-scale restructuring of its own artificial intelligence direction.
Both models support a context window of up to 1 million tokens. This allows processing large documents, analyzing significant volumes of program code, and retaining information about complex tasks during long-term work.
Meta positions Muse Wonder as a universal model for everyday and professional tasks, while Muse Wonder Full is designed for the most complex scenarios, in particular long-term autonomous programming and working with large software projects.
Competition with OpenAI and Anthropic
According to the results of tests published by Meta, Muse Wonder Full surpasses a number of leading competitors in certain programming and AI agent tasks.
In the DeepSWE 1.1 test, which evaluates the ability of artificial intelligence to perform complex software tasks, Muse Wonder Full scored 79 points against 74.1 for GPT-6 Astra and 74.2 for Claude Opus 5.5.
In the SWE-Marathon test, which checks work on long-term software projects, Meta's model received 57.5 points against 42.5 for GPT-6 Astra and 46.3 for Claude Opus 5.5.
At the same time, the new models do not have an advantage in all categories. In particular, competitors demonstrate better results in certain terminal work tests, medical queries, and other tasks. Therefore, the results of individual tests do not mean absolute leadership of Muse Wonder on the market.
Meta notes that some evaluations were conducted by independent organizations, and some by the company's own methods.
AI can independently create applications and fix errors
One of the main features of Muse Wonder has become the ability to perform long-term tasks consisting of many sequential operations.
The model can independently plan work, distribute it among auxiliary AI agents, run tools, check the obtained results, and fix errors.
According to the developers, the new systems are capable of creating full-fledged applications from an initial description, working with large code repositories, and automating tasks in browsers and computer programs.
Besides programming, Meta announces the possibility of using the models for analyzing financial documents, legal contracts, insurance claims, and healthcare-related information.
For this, the company has developed two technologies. Active Memory allows AI to store important information between conversations and return to it when needed. Active Focus helps the model continue performing a complex task, change the approach in case of failure, and recover after errors.
At the same time, the system can ask the user for clarifications and request confirmation before performing important actions.
Availability of new models
Muse Wonder is already available to developers through the Meta Model API and the Muse Code platform. The cost of use via API is $4 per million input tokens and $20 per million output tokens.
The more powerful Muse Wonder Full is currently available only through a paid Muse Code subscription. Access can be obtained in the Power Usage plan costing $50 per month and the Maximum Usage plan for $200 per month.
In the future, Meta plans to integrate Muse Wonder into its personal AI agent Muse. First, subscribers of the service will receive new capabilities, after which the company will develop specialized versions for a wider audience.
The launch of Muse Wonder has become an important stage in Meta's strategy to return to competition in the segment of advanced AI models. Over the past 14 months, the company has been restructuring its research direction, investing in computing infrastructure and developing systems that can not only answer questions but also independently perform practical tasks.
Based on materials from: Meta, Meta AI Developers, The Verge