Meta launches Muse Code, AI that programs itself and challenges OpenAI

Presented this Wednesday (5), Muse Code works directly through the computer terminal and uses the new Muse Spark 1.2 model. The tool is initially available in beta version.

What is Muse Code?

Muse Code is an artificial intelligence agent aimed at software development. Unlike tools that only suggest small snippets of code, the system was designed to follow long projects and execute several steps without requiring constant commands from the user.

AI can analyze a request, organize the project, break the work into smaller tasks, write the code, and verify the results.

Mark Zuckerberg stated that the system can also distribute parts of a complex activity among several “subagents”, who work simultaneously. According to the executive, this structure seeks to increase the speed and quality of responses.

How does Meta's programming AI work?

Meta trained Muse Code and Muse Spark 1.2 together so that the tool and model operated seamlessly.

The system is capable of writing and fixing code, handling long-running development tasks, and verifying that the output produced works as expected.

The agent also keeps a permanent record of the actions taken. If a failure or interruption occurs, he can resume work where he left off, without having to restart the entire project.

How much does it cost to use Muse Code?

Access will be charged according to consumption, and not by a fixed monthly subscription.

In the standard mode, the announced price is US$1.25 per million input tokens, which correspond to information sent by the user, and US$4.25 per million output tokens, related to content produced by artificial intelligence.

The pricing strategy puts Meta in direct competition with companies that already offer tools aimed at programmers and technology organizations.

Meta tries to turn AI investments into revenue

Programming has become one of the most sought-after segments of generative artificial intelligence. Companies and developers already pay for tools capable of creating functions, finding errors, testing systems and performing software engineering tasks.

Meta's entry into this market represents an attempt to generate direct revenue from its AI models. Currently, the majority of the company's revenue, which owns Facebook, Instagram and WhatsApp, still comes from advertising.

The launch occurs at a time of pressure for the company to show a return on billion-dollar investments made in infrastructure, data centers, advanced models and hiring specialists.

What is Muse Spark 1.2?

Muse Spark 1.2 is the new version of the Meta artificial intelligence model used by Muse Code.

The previous generation, Muse Spark 1.1, had been presented in July as a multimodal model focused on reasoning, use of tools, programming and execution of tasks by autonomous agents.

According to Meta, complex challenges generated and evaluated during testing of the previous version helped improve Muse Spark 1.2's ability to follow instructions and execute extensive projects.

Muse Code competes with OpenAI and Anthropic in the market

Meta arrives in a sector that already brings together consolidated products. Microsoft made space for this type of technology with GitHub Copilot, while OpenAI and Anthropic advanced with agents capable of working more independently.

Agentic tools can be given a broader objective and chain different actions to achieve it, such as querying files, changing code, running tests, identifying failures and presenting a final solution.

With Muse Code, Meta seeks to compete in this market by offering integration with its own model, processing tasks in parallel and charging proportional to usage.

ChatGPT launches advertising in Brazil and can use conversations to target ads

Advertising will be restricted to adults, and OpenAI says it does not share personal data with advertisers. Executives in the advertising market say that the tool offers an unprecedented form of segmentation