Meta Launches Muse Code to Challenge Claude Code and OpenAI Codex

Quick Facts

• Meta launched Muse Code, a terminal-based AI coding agent in beta for macOS and Linux, powered by the new Muse Spark 1.2 model.
• A contributor tier offers pricing as low as $0.10 per million input tokens in exchange for permission to train future Meta models on your data.
• On the Terminal-Bench 2.1 benchmark, Muse Spark 1.2 scored 82.9%, trailing Anthropic’s Opus 5 in Claude Code at 86.7% but beating OpenAI’s GPT-5.6 Terra in Codex at 81.8%.

Meta entered the AI coding agent market Wednesday with Muse Code, a terminal-based tool now in beta for macOS and Linux. The release puts the company in direct competition with Anthropic’s Claude Code and OpenAI’s Codex.

CEO Mark Zuckerberg announced the tool in a social media post, describing it as a coding agent that handles complete software engineering tasks. “It’s a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results,” he wrote.

How It Works

Muse Code installs with a single command and runs on top of Muse Spark 1.2, a new version of Meta’s coding model developed and trained alongside the agent. For large projects, it spawns sub-agents that work in parallel in isolated worktrees.

The tool runs persistent background agents throughout each session rather than spawning new ones for individual tasks. Meta says this reduces latency on difficult, multi-step work. Every model call, tool run, and edit is written to a local event log before execution, allowing the agent to resume exactly where it stopped after a crash, even hours into a long task.

Built-in commands include /plan, which breaks a task into an approval-gated plan; /grill, which stress-tests that plan; and /goal, which drives the agent toward a stated objective. Unlike Claude Code and ChatGPT Codex, Muse Code currently has no dedicated app interface.

Pricing and Data Tradeoffs

Meta offers two pricing tiers. The standard pay-as-you-go rate matches Muse Spark 1.1 pricing: $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier drops pricing to $0.10 per million input tokens and $0.20 per million output tokens, roughly 12 to 21 times cheaper, in exchange for explicit permission to use prompts and completions to train future Meta models.

Chief AI Officer Alexandr Wang said Meta is also accepting zero data retention requests from enterprise customers who do not want their data used for model training. “That’s a big enterprise feature that’s really important for a lot of customers,” Wang said. The commitment is notable given that Meta derives 98% of its revenue from advertising, a business built on user data.

Benchmark Results

Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1, placing second behind Anthropic’s Opus 5, which leads at 86.7% in Claude Code. OpenAI’s GPT-5.6 Terra in Codex scored 81.8%, and xAI’s Grok 4.5 in Grok Build scored 81.6%.

On the DeepSWE 1.1 benchmark, Muse Code posted 59.3%, compared to 65.0% for Opus 5 and 64.8% for Codex. On Meta’s internal coding benchmark, Muse scored 70.6% against Opus 5’s 79.4%.

Internal Rollout

Meta is requiring thousands of its engineers to use Muse Code weekly. The tool already has 7,000 active internal users who have generated more than 800 fixes that have improved model performance, according to an internal memo reviewed by The Information.

Wang declined to share external adoption figures for the Muse Spark model family but said uptake has been “exciting and strong.” Zuckerberg framed the broader AI push on Meta’s Q2 2026 earnings call as accelerating the company’s core business while opening new enterprise opportunities.

Read more: Meta takes on Anthropic and OpenAI with its first AI coding agent, Muse Code