Meta Ships a Coding Agent, Offers 12x Discount If It Can Read Your Code

Meta, a company that owns four of the largest communication platforms on Earth and would like you to know it is also a frontier AI lab, released its first coding agent on August 5. It is called Muse Code, it runs in the terminal on macOS and Linux, it is powered by the new Muse Spark 1.2 model, and it can be installed by piping a URL into bash, a practice that every security team on the planet has forbidden and every developer on the planet will do before lunch.

The agent itself is, by the accounts of people who’ve used it, legitimately capable. It plans changes, writes code, verifies its own work, and coordinates “persistent async background agents” that stay active through a session — meaning while you review one change, other agents keep working in parallel git worktrees, like a very industrious haunted house. A local event log records every model call, tool use, and approval, so when your laptop dies mid-refactor, the agents simply resume, unbothered, like nothing happened. The machines have achieved what humanity could not: recovering from a crash without a twenty-minute sulk.

In Meta’s showcase demo, the model ran more than 1,000 tool calls over up to 24 hours optimizing GPU kernels. Twenty-four hours. Unsupervised. Improving the very hardware pipeline that trains its successor. This is presented in the launch materials as a benchmark result and not, as it might read to a lay audience, as the opening paragraph of a cautionary novel.

On the benchmarks, Muse Spark 1.2 posts 82.9% on Terminal-Bench 2.1 — a hair ahead of OpenAI’s GPT-5.6 Terra with Codex (81.8%) and xAI’s Grok 4.5 (81.6%), and behind Anthropic’s Opus 5 with Claude Code (86.7%). The industry has thus arrived at a place where four companies separated by a few percentage points on a benchmark named like a rejected Transformers character each describe themselves as the leader, and each is, depending on the column you sort by.

But the genuinely novel move is the pricing. The standard tier costs $1.25 per million input tokens and $4.25 per million output. The “contributor” tier costs roughly a tenth of that — per Forbes, around $0.10 and $0.20 — in exchange for one small thing: Meta gets to train on your prompts and code.

Twelve to twenty-one times cheaper, if it can watch you work.

Forbes frames this as a data-acquisition play, and it’s a shrewd one: developer code is “checkable” data — it compiles or it doesn’t, tests pass or they don’t — which makes it premium training material. Meta has effectively opened a farmers market where you are both the customer and the produce. The discount isn’t a discount. It’s a salary, and it is not large.

The corporate context explains the urgency. Muse Code comes out of Meta Superintelligence Labs, the division assembled around Alexandr Wang after the $14.3 billion Scale AI acquihire, and it arrives while Meta’s capital expenditure guidance sits in the $130–145 billion range and quarterly free cash flow has thinned from $8.5 billion to $784 million. The stock is down more than 20% over twelve months while Google is up 80%. When your infrastructure bill has twelve digits, “developers will pay us to give us training data” is not just a product strategy; it’s a coping mechanism with revenue.

Still — a real agent, really shipped, priced aggressively, with an honest-to-goodness novel trade on the table. The terminal-only, no-GUI stance is almost charmingly austere, a coding agent for people who think buttons are a moral failing.

Every developer now faces the same quiet arithmetic at the install prompt: my code is worth a 12x discount to them; what is it worth to me? An entire industry is about to answer that question one curl | bash at a time, and Meta is betting — $134 billion or so — that the answer is “eh, take it.”

Sources: