Meta's New Coding Agent: 21x Cheaper for Your Code
On August 5, 2026, Mark Zuckerberg personally announced Meta’s first AI coding agent, Muse Code, running on a new coding-specific model called Muse Spark 1.2. This is a company arriving a full year late to a category Anthropic and OpenAI have already been selling — and the way it’s chosen to fight back is by offering a price deal neither of those competitors has matched. The catch is written on the back of that deal: you have to hand over the code you write.
What Muse Code actually is
Muse Code works the way Claude Code, Codex CLI, and Grok CLI do: you type what you need into a terminal, and it goes and writes or edits code directly in your project. It’s powered by Muse Spark 1.2, which has a 1-million-token context window and was co-trained specifically with the Muse Code harness. It can automatically fan a single job out across multiple parallel subagents, each working in its own isolated git worktree so they don’t step on each other. Every step gets logged as a JSONL event trail. Meta’s own description: “Every subagent it spawns, every tool call, every steer and cancel, is observable and replayable through the event log.”
The built-in commands worth knowing
Muse Code ships with a handful of notable commands: /plan grounds its strategy in the actual codebase and saves its decisions to a .agents/plans/ folder, requiring human approval before execution; /grilling validates each decision through a sequence of targeted questions; /grill-with-docs captures that validation process as a durable decision record inside the project’s own documentation; and /taste is a visual quality filter meant to strip out the look that instantly reads as machine-generated. Installation is a one-line terminal command: curl -fsSL https://dev.meta.ai/install.sh | bash. It’s currently available through Muse Code beta, Meta’s own Model API, and OpenRouter.
Where it actually ranks on performance
On DeepSWE, the benchmark most relevant to coding work, Muse Spark 1.2 scored 59.3% — just below Opus 5 and GPT 5.6 Turbo. For reference, Qwen 3.8 Max scored 56.6% on the same test. Notably, the current category leaders, GPT 5.6 Sol and Fable 5, weren’t included in Meta’s own comparison charts. Meta’s own claim is that Muse Spark 1.2 shows “moderate improvement” over the prior Muse Spark 1.1 across TerminalBench, DeepSWE, its internal code bench, and GDPVAL. In other words: for developers already using a higher-ranked model, Muse Code’s underlying model currently doesn’t beat the strongest options in its own category.
The real story is the pricing
Muse Code offers two pricing tiers with identical functionality: Meta’s own announcement specifies the standard tier at $1.25 per million input tokens and $4.25 per million output tokens, while its description of the “Contributor” tier only says it’s billed by rolling 5-hour token volume rather than by request count, without publishing an exact rate. Forbes, however, reports the Contributor tier’s actual rate as just $0.10 per million input tokens and $0.20 per million output — 12.5x to 21x cheaper than standard. The difference is what you’re agreeing to: pick the cheap Contributor tier, and you’re consenting to let Meta use the prompts and completions you send through it to train its own models. One commentator put it bluntly: this is effectively “Meta posting an offer to buy your source code, and the currency is compute.” Code is especially valuable training material for Meta because its output can be checked directly against compilation and tests — unlike most business processes, where feedback takes weeks to arrive, if it arrives at all. If you don’t want your code used for training, Meta does offer a zero-data-retention option, but only by contacting its sales team directly.
Why Meta is racing to grab share right now
This pricing push is landing at a moment of real financial pressure for Meta: the company posted $60.8 billion in Q2 2026 revenue, beating estimates, but free cash flow shrank sharply to just $784 million from $8.5 billion a year earlier, even as its 2026 capital spending guidance still sits at $130-145 billion. Its stock dropped roughly 8% after the earnings report. Against that financial backdrop, trading a steep discount for developers who voluntarily contribute high-quality, verifiable code as training data is, in a way, a good deal for Meta — it’s just that the other side of that deal is every developer who hits “Contributor tier” to save money.
