Meta announced an early beta of Muse Code on August 5, 2026, a terminal-based coding agent built to compete with Anthropic's Claude Code and OpenAI's Codex (Engadget, CNBC). The agent runs on Muse Spark 1.2, a new version of Meta's main AI model tuned specifically for software development tasks (Spokesman-Review, Meta Research).

Muse Code can write code, plan changes across a project, and validate that its output works correctly, meaning you can describe a feature in plain text and get working software back. It can also spin up sub-agents (additional instances of itself that handle smaller pieces of a task in parallel) and delegate work to them, which lets it complete larger jobs more efficiently (Engadget). Muse Spark 1.2 itself brings improvements in code generation, complex debugging, understanding of existing codebases, and end-to-end developer workflows (Meta Research).

Where Muse Code stands out is pricing. By default, it uses the same pay-as-you-go rates as the Muse Spark model: $1.25 per million input tokens (the text you send to the model) and $4.25 per million output tokens (the text it generates back). Meta Chief AI Officer Alexander Wang said the company will also offer a contributor tier at a much lower cost. That tier requires users to agree to provide feedback that helps improve the agent, and it costs $0.10 per million input tokens and $0.20 per million output tokens (Engadget, WSJ).

For comparison, Anthropic's Sonnet 5 model typically costs $3 per million input tokens and $15 per million output tokens (Engadget). The contributor tier undercuts those rates by a factor of 30 on input tokens and 75 on output tokens, though the feedback obligation means users are effectively trading their interaction data for the cost savings.

Muse Spark is a natively multimodal reasoning model, meaning it can process and generate different types of data, not just text. It supports tool use, visual chain of thought (showing its reasoning steps through visual representations), and multi-agent orchestration. The model was first introduced in April 2026 (Meta AI Blog). An intermediate version, Muse Spark 1.1, arrived in July 2026 with gains in tool and computer use, coding, and multimodal capabilities (Meta AI Blog). Meta evaluates the Muse Spark models across benchmarks covering reasoning, multimodal performance, coding, tool use, and health knowledge (Meta AI Evaluation Methodology). According to Meta's Muse Spark Safety and Preparedness Report, published in May 2026, the model generates secure code at rates competitive with peer models (Meta AI Safety Report).

The Muse family extends beyond code. Muse Spark and Muse Image work together to combine code and media generation, enabling the creation of animated GIFs and websites with embedded media (Meta AI Blog). Muse Image follows instructions faithfully, edits with precision, composes from multiple references, and draws on Instagram for social context. Muse Video delivers visual fidelity with native audio support (Meta AI Blog).

The broader context here is that Meta is entering the coding agent market with a pricing strategy that, if sustained, applies real margin pressure on Anthropic and OpenAI. The contributor tier in particular is aggressive: sub-cent pricing per million input tokens is unusual for frontier-class models, and the feedback-for-discount exchange gives Meta a direct pipeline of developer interaction data to refine the agent. Whether that data advantage turns into a quality advantage will depend on how much useful signal Muse Code can extract from contributor sessions versus what Anthropic and OpenAI already collect through their own developer ecosystems.

Meta is also distinguishing itself by shipping a terminal-based agent, the same form factor Claude Code adopted. This is a deliberate choice for an audience that already works in the terminal and wants an agent that operates close to the filesystem and build tools rather than through a browser-based chat window. The sub-agent orchestration capability positions Muse Code not just as a code completion tool but as a system that can break down tasks, parallelize work, and validate its own output.

The integration story matters too. If Muse Code can reliably orchestrate Muse Image and Muse Video alongside code generation, Meta is offering something its competitors do not currently match: a single provider stack for generating code, images, video, and audio from a unified model family. The practical value of that integration at this early stage is unproven, but the architectural direction is clear.

Meta is betting that aggressive pricing, multimodal integration, and a contributor feedback loop will be enough to pull developer attention away from established competitors. The beta will tell us how close Muse Code is to that ambition.