Meta launches Muse Code AI agent to challenge OpenAI and Anthropic

Meta is aggressively expanding its footprint in the artificial intelligence sector even as investors scrutinize heavy capital expenditures and a soft financial outlook. Under pressure to diversify its income beyond a core advertising business that generates roughly 98% of its revenue, the company has introduced its first dedicated software development tool to challenge established market leaders (according to International Business Times reporting).

The newly unveiled coding agent, Muse Code, enters an increasingly crowded market dominated by rival tools such as Anthropic's Claude Code and OpenAI Codex. Rather than claiming superior performance over these established alternatives, Meta is leading with a aggressive pricing strategy and deep cost discounts to attract developers (as noted by Morningstar). The financial push arrives at a critical juncture for the firm, which reported declining second-quarter free cash flow amid soaring infrastructure costs for data centers and computing capacity.

Related YouTube video

Developed under the Meta Superintelligence Labs division led by AI chief Alexandr Wang, Muse Code is powered by the freshly debuted Muse Spark 1.2 model. The foundation model and coding agent were trained together to optimize software engineering capabilities, including complex debugging and complete repository management. Operating on macOS and Linux through a terminal harness, the agent deploys persistent subagents that remain active throughout an entire session rather than spawning separately for isolated prompts. This architecture reduces latency and repeated information gathering during multi-stage projects.

Under the hood, the runtime records every model request, tool execution, approval, and code edit in a local event log. This mechanism enables developers to resume interrupted sessions or recover exact system states following a crash. Built-in skills guide the workflow: the /plan command translates tasks into user-approved roadmaps, the /grill command probes those plans for potential vulnerabilities, and /goal directs the system toward completion. In demonstration examples provided by Meta, the agent successfully interpreted a home video tour to build a functional vacation property marketing and booking website, and executed over 1,000 tool calls across 24-hour sessions to optimize GPU kernels for Nvidia Hopper hardware.

To undercut competitors, Meta has established a transparent pricing schedule for the preview release. Users can access the tool via a pay-as-you-go model priced at $1.25 per million input tokens and $4.25 per million output tokens (reported by Newsbytesapp). Furthermore, the platform incorporates a contributor tier described by Wang as being more than 10 times cheaper than the standard pay-as-you-go rate. Alongside direct availability through the Meta developer portal and API, Muse Spark 1.2 is slated for distribution on OpenRouter alongside models from labs such as DeepSeek and Z.ai.

To ease corporate security anxieties surrounding proprietary intellectual property, Meta is rolling out a zero-data retention policy for enterprise clients (noted by International Business Times). By accepting requests to omit customer code from future training sets, the company aims to court corporate users handling sensitive internal codebases.

While Wang declined to share specific user metrics for the Muse Spark family, he maintained that initial customer response has been encouraging, stating that adoption has been exciting and strong.

Expanding the Developer Ecosystem

According to International Business Times reporting, the newly unveiled coding agent was introduced Wednesday by Meta Superintelligence Labs. Cryptobriefing notes that the software operates in beta on macOS and Linux, utilizing persistent subagents that remain active across an entire session. Morningstar points out that the offering arrives as the company faces investor pressure to generate tangible income from infrastructure investments.

As the competitive race in developer tooling intensifies, the company intends to roll out additional features and more capable models. The next step involves expanding global access and deploying further model iterations through the Meta developer portal and partner networks.

Training and Architecture of the Muse Code Platform

According to Cryptobriefing, the Muse Spark 1.2 model received increased training compute dedicated specifically to coding tasks alongside an expansion of the environments used during the training process. The same reporting notes that Meta Superintelligence Labs trained the model concurrently with the coding agent to refine how the system handles tool invocation, context management, and planning structures. During testing phases highlighted by Cryptobriefing, the foundation model executed over 1,000 tool calls spanning 24-hour sessions to optimize GPU kernels for Nvidia Hopper hardware, resulting in measurable performance gains over baseline implementations.

International Business Times and Newsbytesapp report that Alexandr Wang leads Meta Superintelligence Labs, the division responsible for developing the platform. While Wang declined to disclose specific adoption statistics for the Muse Spark family, he stated in an interview with International Business Times that customer interest has been strong and encouraging. The software arrives as CEO Mark Zuckerberg seeks to establish Meta as a dominant player in advanced software development while developing revenue streams beyond its advertising business, which currently accounts for roughly 98% of total revenue.

As competition among developer platforms intensifies, the next step involves rolling out additional features and deploying more capable models through the Meta developer portal, API access points, and partner distribution networks.