Meta announced an early test version of Muse Code on August 5, 2026. It is an AI tool that helps people write software by understanding plain-language instructions and producing working code. The tool is designed to compete with similar products from Anthropic and OpenAI (Engadget, CNBC). Muse Code is powered by Muse Spark 1.2, a new version of Meta's main AI model that has been specially tuned for programming tasks (Spokesman-Review, Meta Research).

Muse Code can write code, plan out changes to a software project, and check that its work actually functions. You describe what you want in everyday language, and the tool produces the software to match. It can also create helper copies of itself to work on smaller pieces of a big task at the same time, which speeds up larger jobs (Engadget). Muse Spark 1.2 brings improvements in writing code, fixing difficult bugs, understanding existing projects, and supporting the full workflow a developer goes through (Meta Research).

The biggest surprise is the price. AI models charge based on how much text you send them and how much they send back, measured in tokens, which are roughly pieces of words. By default, Muse Code costs $1.25 per million tokens you send in and $4.25 per million tokens you get back. Meta Chief AI Officer Alexander Wang said the company will also offer a much cheaper option called the contributor tier. Anyone who agrees to provide feedback to help improve the tool pays just $0.10 per million input tokens and $0.20 per million output tokens (Engadget, WSJ).

For comparison, Anthropic's Sonnet 5 model typically costs $3 per million input tokens and $15 per million output tokens (Engadget). The contributor tier undercuts those prices by a factor of 30 on input and 75 on output. The trade-off is that users are effectively handing over their usage data in exchange for the discount.

Muse Spark is a multimodal model, which means it can work with different kinds of content, not just text. It can use external tools, show its reasoning process visually, and coordinate multiple agents working together. The model was first introduced in April 2026 (Meta AI Blog). An updated version, Muse Spark 1.1, arrived in July 2026 with improvements in tool use, coding, and handling multiple types of content (Meta AI Blog). Meta tests the models across areas including reasoning, coding, tool use, and health knowledge (Meta AI Evaluation Methodology). According to Meta's safety report published in May 2026, the model produces secure code at rates comparable to similar models from competitors (Meta AI Safety Report).

The Muse family goes beyond code. Muse Spark and Muse Image can work together to combine code and media, making it possible to create animated GIFs and websites with embedded images and video (Meta AI Blog). Muse Image follows instructions precisely, edits accurately, and pulls social context from Instagram. Muse Video produces high-quality visuals with sound (Meta AI Blog).

The broader context is that Meta is entering the coding tool market with a pricing strategy that, if it holds, puts real pressure on Anthropic and OpenAI. The contributor tier is especially aggressive: charging less than a cent per million tokens sent in is unusual for top-tier AI models. The feedback-for-discount arrangement also gives Meta a direct stream of data showing how developers actually use the tool, which could help Meta improve it faster than competitors. Whether that data advantage leads to a better product will depend on how much useful information Meta can pull from those sessions versus what Anthropic and OpenAI already gather from their own users.

Meta is also making a deliberate choice by building a tool that runs in the terminal, the text-based interface that many programmers already use daily. This is the same approach Claude Code took. It targets developers who want an AI assistant that works directly with their files and tools rather than through a web browser. The ability to manage helper agents positions Muse Code as more than a tool that finishes your sentences. It is a system that can break a task into parts, work on them in parallel, and check its own results.

The integration between Muse Code, Muse Image, and Muse Video also matters. If it works reliably, Meta would be offering something no competitor currently matches: one company's tools for generating code, images, video, and audio all from the same family of AI models. How useful that integration actually is at this early stage is unproven, but the direction is clear.

Meta is betting that low prices, the ability to handle multiple types of content, and a feedback loop with users will be enough to draw developers away from the competitors they already use. The beta test will show how close Muse Code is to that goal.