Meta Launches Muse Code Coding Agent, Undercuts Anthropic on Price by Up to 98%

Meta has entered the AI coding agent market with Muse Code, a terminal-based tool now available in early beta. The product is powered by Muse Spark 1.2, a new version of Meta's proprietary model, and directly targets Anthropic's Claude Code and OpenAI's Codex.

What Muse Code does

According to Meta, Muse Spark 1.2 delivers "improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows." Muse Code can write code, plan multi-step changes, and validate results — turning text prompts into working software. The agent can also spin up and coordinate multiple sub-agents, delegating tasks to speed up larger jobs.

Meta's launch materials include sample projects built with the tool: an interactive simulation of a photon sphere and a clone of Plants vs. Zombies. A demo video shows the agent generating a webpage directly from an MP4 file input.

Pricing is the differentiator

Meta is pricing Muse Code aggressively. The default pay-as-you-go rate matches Muse Spark's standalone API pricing: $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.

Meta Chief AI Officer Alexander Wang told CNBC the company will also offer a "contributor tier" at a steep discount. According to the Wall Street Journal, that tier costs $0.10 per 1 million input tokens and $0.20 per 1 million output tokens — but requires users to agree to provide usage feedback that Meta can use to improve the model.

For comparison, Anthropic's Sonnet 5 typically costs $3 per 1 million input tokens and $15 per 1 million output tokens. At the contributor tier, Muse Code's output pricing is roughly 98.7% cheaper than Sonnet 5's; even at the standard rate, output tokens cost about 72% less.

Meta has not disclosed a context window size for Muse Spark 1.2, nor published independent benchmark scores for code generation or debugging accuracy. All performance claims in this launch come directly from Meta and have not been independently verified.

What this means

Meta is betting that price, not just capability, will drive adoption of coding agents. The strategy mirrors what happened when Chinese labs like DeepSeek undercut US model pricing — several companies migrated workloads specifically to cut costs, even when raw performance lagged behind frontier models from OpenAI and Anthropic.

Whether that logic holds for Muse Code depends entirely on real-world performance, which remains unverified. Coding agents are judged on reliability across long, multi-step tasks — debugging, test validation, and codebase navigation — not just raw token cost. If Muse Spark 1.2 underperforms Claude Code or Codex on those fronts, the price gap won't matter to engineering teams that value accuracy over savings.

The contributor tier is also notable: Meta is effectively paying users in the form of discounted pricing to generate the feedback data it needs to improve Muse Spark. That's a direct trade of cost savings for training signal, and it suggests Meta sees the current model as a work in progress rather than a finished competitor to established coding tools.