Two new models bring Meta deeper into paid coding tools and AI agents, with the strongest version restricted to higher subscription tiers.
Meta Muse Wonder and Muse Wonder Full launched on October 12, giving the company a new pair of advanced AI models aimed at coding, reasoning and tasks that require sustained work across software. Meta calls them its most capable models yet. Muse Wonder is available through the Meta Model API and Muse Code, while Wonder Full has a narrower subscription route.
The release puts Meta more directly into competition with OpenAI and Anthropic, Reuters reports. For users, the immediate questions are practical: which model can they access, what does it cost, and how well does it perform on the work they need done? A claim of competitive benchmark performance does not settle all three.
What Meta Muse Wonder is designed to do
Meta describes two central capabilities: Active Memory and Active Focus. Active Memory is intended to retrieve relevant information from earlier conversations. Active Focus describes persistence through difficult tasks, including trying alternative approaches, recovering from errors and checking with the user before important actions. These are Meta’s descriptions of the intended behaviour, rather than guarantees that every task will succeed.
The models accept text, images, video, audio and PDFs, and return text. Meta’s developer documentation lists a context window of 1,048,576 tokens, giving the system space to process substantial material within an interaction. A large context window, however, does not itself establish reliable recall or accurate interpretation of every document.
Coding is a prominent use case. Meta positions Wonder for changes spanning multiple files, debugging, testing and longer software projects. That makes the developer experience important: a model must work with repositories and tools, preserve requirements and verify its changes, rather than simply produce a convincing code sample.
Benchmarks show ambition, not an overall verdict
Meta’s evaluation materials compare the Wonder models with systems including Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra. The tests cover software engineering, computer use, memory and professional tasks. They support a claim that Meta is pursuing frontier-level competition, but should not be read as proof that either model is best at every activity.
The methodology combines internally generated results with third-party and provider-reported scores. Meta acknowledges that its competitor runs may not use optimally tuned prompts, tools or runtimes. Some coding comparisons also use different agent products or test frameworks, which limits what a single ranking can establish.
For buyers, the relevant test is repeatable performance on their own tasks. A developer fixing production bugs, a researcher comparing documents and a business preparing customer communications may reach different conclusions. Cost, correction time and the ability to recognise uncertainty belong alongside benchmark scores in that assessment.
Access and pricing are different for each model
Wonder Full is not available through the public Meta Model API. Meta lists it for Muse Code’s Power Usage and Maximum Usage plans, advertised at $50 and $200 per month respectively. The regular Wonder model is available through the API and Muse Code. Choosing maximum reasoning effort on regular Wonder does not turn it into Full.
For API use, Meta prices Wonder at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. These are usage charges, separate from a coding subscription. Repeated agent steps can accumulate additional input and output, so the cost of completing a task matters more than the price of one response.
Muse Code’s documentation says subscription benefits and availability may vary by region. Developers in Ghana and other African markets should check the entitlement shown during onboarding before purchasing. The global announcement does not establish identical plans, payment options or access in every country.
Muse integration and a privacy choice
Muse is Meta’s personal-agent product; Muse Code is its coding tool; Wonder is a model that can power those experiences. Meta says Wonder will reach more Muse experiences soon, starting with subscribers. The announcement does not provide a firm rollout date or establish immediate access for all consumer users.
There is also a consequential subscription setting. Muse Code defaults to Contributor mode, which provides more usage in exchange for permission to use submitted content to train future Meta models. Subscribers can disable it. Meta says Standard mode does not use that content for training. Teams working with confidential code or documents should review the setting before uploading material.
Analysis: Distribution opens doors, usefulness decides
The launch builds on Muse Spark and follows the June 2025 move of Scale AI founder Alexandr Wang to Meta. Those milestones establish a development timeline; they do not prove that leadership changes alone produced the models’ capabilities.
Meta reported 3.60 billion average daily active people across its family of apps in June 2026. That is a substantial potential distribution advantage, but it is not a count of Muse users, paying subscribers or people receiving Wonder.
Meta now has another product to offer developers and agent users. Its next test is whether the models deliver dependable results at a cost people accept, with access and data controls they understand. Launch benchmarks can attract attention; sustained use will reveal whether Wonder earns a place in everyday work.
