Meta Challenges OpenAI with Muse Code, Its Economic Agent

An agent that distributes work among multiple sub-agents

Muse Code, Meta's programming tool, stands out for its operation exclusively in the terminal, without a graphical interface. It is designed to handle complete software engineering tasks across entire repositories. The tool plans the necessary modifications, writes the code, and then checks the results obtained. To manage complex tasks, Muse Code employs several sub-agents that work in parallel, each in an isolated workspace, ensuring that the user's working copy remains intact. During testing, this tool demonstrated its ability to simultaneously develop six features of a game without causing conflicts.

Meta has also integrated background agents that remain active throughout the session. This helps avoid the repetition of information searches that have already been conducted. An event log continuously records every model call, every tool usage, and every file modification. In the event of a crash, the agent can resume its work exactly where it left off, which is particularly useful for long tasks that may last several hours.

Three main skills are provided by default with Muse Code:

• /plan: this command transforms a request into a detailed, step-by-step validated action plan.
• /grill: it tests the robustness of the plan by identifying potential flaws.
• /goal: it guides the agent towards achieving a specific goal until completion.

Muse Spark 1.2, a model that progresses without dominating benchmarks

Muse Code is powered by the Muse Spark 1.2 model, an update of Spark 1.1 specifically trained for code. This model has been co-trained with Muse Code to optimize their functioning together. On the benchmarks published by Meta, Muse Spark 1.2 generally ranks behind Anthropic's Claude Opus 5 and Codex's GPT-5.6 Terra. While the gap is moderate on benchmark tests like Terminal-Bench, it widens on longer and more complex exercises, where Muse Spark 1.2 finishes several points behind its competitors.

However, Muse Spark 1.2 stands out on the MCP Atlas test, dedicated to generalist agents, where it surpasses all its rivals, including Claude Opus 5. In contrast, on the GDPVal-AA test, Opus 5 maintains its lead. Although Muse Spark 1.2 manages to compete with its rivals, it does not significantly stand out.

A beta accessible with a single command, billed by usage

Muse Code is now available in beta for macOS and Linux systems. Installation is done simply with a single command line in the terminal. The tool also serves as an orchestration layer capable of driving third-party models, although Meta admits that the experience is optimized for Muse Spark 1.2, as both products were developed together.

Muse Spark 1.2 is also accessible via the Meta Model API, with expanded international access compared to the previous version. Companies handling sensitive code can request that their data not be retained.

To compete with Anthropic and OpenAI, Meta bets on pricing

Muse Code represents Meta's first programming agent, which had already taken a step in 2023 with Code Llama, an open-source model without an associated autonomous agent. Meta thus competes with Anthropic's Claude Code and OpenAI's Codex, both already well-established in a key segment for generative AI revenue, without showing a clear superiority in benchmarks.

To stand out, Meta relies on its ability to drive third-party models, unlike Claude Code and Codex, each tied to its own family of models. Most importantly, Meta offers attractive pricing: Muse Code charges $1.25 per million input tokens and $4.25 for output, compared to $5 and $25 for Claude Opus 5 via Anthropic's API, and $2 and $12 for GPT-5.6 Terra at OpenAI. A "Contributor" offer, described by Alexandr Wang as ten times cheaper, is also available for users willing to share their usage data to improve the model.