A Fourth Contender in the Coding Wars — Meta's 'Muse Code' Cuts Prices to a Tenth or Less If You Hand Over Your Data
Meta releases its first coding agent, "Muse Code," in beta. Terminal-based and powered by Muse Spark 1.2, it drops to a tenth or less of competitors' prices if you consent to sharing your data — but the discount is also the price of handing over your code.
On August 5, 2026, Meta released its first coding agent, "Muse Code," in beta. Mark Zuckerberg described it as "a terminal-based agent that handles end-to-end software development work spanning large repositories." With Claude Code (Anthropic), Codex (OpenAI), and Cursor all vying for position, the coding agent market now has one more heavyweight in the ring.
A Meta-Branded Agent Shows Up in the Terminal
Muse Code targets macOS and Linux, and installation starts with a single command. There's no dedicated GUI app of the kind ChatGPT's Codex or Claude Code offer — the whole thing is built around the terminal alone. Powering it is a new model, "Muse Spark 1.2." Meta says it is "better at using tools, requires fewer retries, and produces higher-quality output than approaches that simply wrap an external skin around a general-purpose model," citing the fact that it was co-trained alongside the agent as a key strength. The context length is 1 million tokens.
Designed to Run Multiple Agents in the Background
The design centers on running agents in parallel with divided roles, rather than having a single model answer one question at a time. Its notable features include the following.
- Persistent asynchronous agents: line up multiple implementation workers while a reviewer runs in the background.
- Local logging: model calls, tool usage, approvals, and edits are recorded locally so you can trace what happened.
- Crash recovery: if it goes down, you can pick up the work where it left off.
- Commands for planning and stress testing: run planning and stress tests through built-in commands.
Two-Tier Pricing, With a "Data for Discount" Option
The pricing is what catches the eye. Even the standard rate undercuts competitors, but if you opt into the "Contributor" tier by agreeing to provide feedback, it drops further — to a tenth or less. On that tier, however, the data you input may be used to improve the model. For reference, Anthropic's Sonnet 5 costs $3 for input and $15 for output (both per million tokens).
| Tier | Input (per 1M tokens) | Output (per 1M tokens) | Notes |
|---|---|---|---|
| Standard | $1.25 | $4.25 | Cached input $0.15 / prompts not used for training |
| Contributor | $0.10 | $0.20 | Cached input $0.002 / data may be used to improve the model |
| (Reference) Sonnet 5 | $3 | $15 | — |
It's available through the Meta Model API (public preview) and via OpenRouter.
What Changes for Developers
Coding agents have already reached the point where the cost of the model directly hits your operating costs. The reason cheap Chinese open models have made inroads into U.S. development teams is precisely this sensitivity to the "cost of keeping it running." Muse Code's standard rate comes in under Sonnet 5, and factoring in the Contributor tier widens the available price range by another notch. Another major player entering the field — and at a low price at that — looks likely to push model selection back from "performance is the only criterion" toward a tug-of-war between performance and cost. And because it's usable via the API and OpenRouter, it's easy to slot into existing workflows and try out, which should help adoption.
What to Check Before Jumping In
The low price and new features are appealing, but there's also reason not to take them at face value. First, the published benchmarks (Terminal-Bench, DeepSWE, Meta's internal coding benchmark, GDPval) are presented only as images, with no accompanying explanation of the measurement methodology — as critics have pointed out. The numbers warrant some reservation. Second, the Contributor tier that cuts the price to a tenth or less is also the price of "handing over your own code and prompts." For teams dealing with highly confidential code, it's worth carefully weighing what leaves your hands in exchange for the discount. Given that it's still in beta, the safe move is to start with evaluation use cases.
References: 9to5Mac / Engadget / Unite.AI / VentureBeat / CNBC
