![\"Photo](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/08\/1786029864_561_0x0.jpg\")
<\/p>\n<\/p>\n
The Meta AI logo is displayed on a smartphone screen placed on a reflective surface onto which the Instagram logo is projected, in Creteil, France, on July 9, 2026. Meta announces the launch of Muse Image, its first proprietary image generation model. (Photo by Samuel Boivin\/NurPhoto via Getty Images)<\/p>\n
NurPhoto via Getty Images<\/p>\n
Meta shipped its first coding agent on Wednesday. Muse Code<\/a> installs from the terminal with a single command and takes on whole engineering jobs across large repositories, planning the change, writing the code, and checking the result. Several agents work a task at once<\/a>, with implementation running in parallel while reviewers watch in the background. It runs on Muse Spark 1.2, a coding-focused model Meta co-trained with the agent<\/a> so the two fit together. The product is in beta.<\/p>\n The launch reads as Meta arriving late<\/a> to a category Anthropic and OpenAI have been selling for a year. Then you get to the price list, which has two columns.<\/p>\n Two Prices For The Same Model<\/p>\n The standard tier costs $1.25 per million input tokens and $4.25 per million output, with cached input at $0.15. That is ordinary pricing for a capable coding model, roughly where the market sits.<\/p>\n The contributor tier costs $0.10 per million input tokens and $0.20 per million output, with cached input at $0.002. Input runs about twelve times cheaper and output about twenty-one times cheaper. The model is identical. The difference is a permission: on the contributor tier, Meta uses your prompts and completions to improve its models. The standard tier does not allow that.<\/p>\n Meta did not price a discount tier. It posted an offer to buy your source code, and the currency is compute.<\/p>\n Why Code Is The Data Worth Buying<\/p>\n Training data has a quality problem that money alone does not solve. The public internet has been scraped, licensing deals are expensive and slow, and synthetic data drifts away from reality when a model learns mostly from itself. What remains scarce is data generated by people doing real work under real constraints.<\/p>\n Code sits at the top of that list for a reason no other domain matches. It is checkable. A program compiles or it does not, tests pass or they fail, and the reviewer either approves the change or sends it back.<\/p>\n Every session on the contributor tier produces a problem, an attempt, and a verdict on whether the attempt worked, inside a real repository with real dependencies and a real bug. Scraped GitHub gives a model the finished commit. A coding agent inside a working developer’s terminal gives it the reasoning, the failures, and the correction that produced the commit.<\/p>\n Meta has run this arrangement before. Facebook’s social graph was built by three billion people handing over their interests in exchange for a service priced at zero, with the terms of service doing the accounting. The company’s advertising business rests on data users produced while doing something else. Muse Code applies the same structure to the most valuable workforce on the internet, and this time Meta is paying developers to take part.<\/p>\n The Spending That Sits Behind It<\/p>\n The timing says something about how much Meta wants this. The company reported second quarter revenue of $60.8 billion in late July, ahead of estimates, while earnings of $6.18 per share came in well under the $7.14 consensus<\/a>. Free cash flow fell from $8.5 billion a year earlier to $784 million. Capital spending guidance for 2026 now runs from $130 billion to $145 billion. The stock dropped about eight percent.<\/p>\n The market read that quarter as spending without a return attached. A tier priced twelve to twenty-one times below standard turns some of that compute into training data the company would otherwise have to license, scrape, or manufacture. Whether it eventually pays is an open question. What the pricing settles is that Meta is treating developer behavior as an asset worth buying at a steep discount to its own list price. That is a decision made at a moment when its cash flow leaves very little room for gestures.<\/p>\n Worth noting on the product itself: Meta published results on Terminal-Bench, DeepSWE, its internal coding benchmark, and GDPval, as images with no methodology write-up<\/a>. Benchmark claims released that way are marketing until somebody independent runs them.<\/p>\n Why Coding And Not Everything Else<\/p>\n