Meta has unveiled a beta version of its first AI coding agent, Muse Code. CNBC reported on Aug. 5 that Meta has entered the AI coding tools market led by Anthropic and OpenAI with Muse Code.

Muse Code helps with app development through a single user interface and supports management of AI-based digital agents used in software development. Like Anthropic's Claude and OpenAI's Codex, it supports a broad range of software engineering tasks, including planning changes, writing code and verifying results.

Muse Code is based on Meta’s latest AI model, Muse Spark 1.2. Like Muse Spark 1.1, unveiled in July, Muse Spark 1.2 is specialised for coding. Meta said the new version was developed and trained together with Muse Code to strengthen coding performance.

Alexandr Wang (알렉산드르 왕), who leads Meta Superintelligence Labs, described Muse Code as a terminal-based coding agent. He said users can install it with a single command and then assign it various software engineering tasks.

Meta also highlighted pricing as a differentiating factor. Developers can use Muse Code on a pay-as-you-go basis, at a price similar to Muse Spark 1.1. At the time, it cost $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.

Muse Code also offers a lower-priced contributor tier. This plan is more than 10 times cheaper than the pay-as-you-go option, but developers must participate in improving the model in an industry-standard way.

Meta also offers Harness, a management framework tuned for coding projects. Developers can switch to other AI models or swap Harness as needed, but Meta said performance is best when its own models and tools are used together.

Meta has also begun accepting requests for zero data retention, under which it does not store developer data for model improvement or other uses. Wang said this is an important feature for corporate customers.