Meta Launches Muse Code, Its First AI Coding Agent, to Challenge Claude Code and Codex

Meta is entering the AI coding assistant race with the beta launch of Muse Code, a terminal-based agent powered by its new Muse Spark 1.2 model, arriving just a month after the Spark 1.1 launch.

The release comes after Meta largely sat out a market dominated by Anthropic’s Claude Code and OpenAI’s Codex.

Built for developers working across large, complex codebases, Muse Code also marks a strategic shift away from Meta’s free Llama approach toward a paid, proprietary developer tool designed to compete with established rivals. 

A Harness Built to Run in Parallel

Meta’s developer blog describes Muse Code as a terminal-based coding agent built for long-horizon software engineering. 

According to Meta, it can handle complex, multi-file changes across large codebases using always-active sub-agents that keep working through sessions lasting hours. 

CEO Mark Zuckerberg, who predicts a future AI-agent boom, outlined the parallel processing approach in a post on X, explaining that large jobs fan out to separate sub-agents working simultaneously in isolated worktrees. 

He added that Meta tested the system by having it build six features for a game at once without any collisions. 

Every model call, tool run, and edit is logged locally, which Meta says lets interrupted tasks resume from where they stopped instead of restarting from scratch, continuing the company’s broader AI infrastructure push. 

Cost Positioned as the Main Selling Point

TechCrunch reported that Meta’s AI Chief Alexander Wang framed Muse Code’s value proposition around affordability as much as raw capability, telling the Wall Street Journal the tool represents a strong option for a lot of workflows and use cases, particularly on cost. 

That pricing strategy fits into Meta’s broader AI spending pattern this year, TechCrunch noted, pointing to the company’s June launch of an enterprise-focused business agent for customer service as part of the same push to expand AI revenue streams beyond its core advertising business. 

Reuters similarly framed the release as Meta directly entering a market where it competes with Anthropic’s Claude models and OpenAI’s Codex amid a broader industry race to monetize AI coding assistants.

This indicates that Meta trained Muse Spark 1.2 and Muse Code together specifically so the model and its harness would work smoothly as a matched pair.

Following a Rocky Stretch for Meta’s AI Ambitions

CNBC said Wednesday’s launch is part of Meta’s effort to rebuild its AI standing after Llama fell short of developer expectations last year, prompting Mark Zuckerberg’s $14.3 billion acquihire of Scale AI and its CEO, Alexandr Wang, who now oversees the company’s AI products. 

While Muse Spark 1.2 isn’t positioned as a bleeding-edge frontier model, Meta still benchmarked it directly against heavyweight systems including Anthropic’s Claude Opus 5, xAI’s Grok 4.5, and similar offerings from OpenAI and Google.

The benchmarks show Meta is confident in Muse Spark 1.2, but the company avoided saying it outperformed competing models. 

Meta’s stock traded flat in after-hours trading following the announcement, suggesting investors are still waiting to see how Muse Code performs against entrenched rivals before rewarding the bet.

Source: Meet Muse Spark 1.2 and Muse Code, the first coding agent from Meta