Meta's New Coding Model Is 92 Percent Cheaper. You Pay With Your Codebase.

Meta shipped a new version of its coding model, Muse Spark 1.3, on September 2, and buried in the release is a pricing move worth more attention than the model itself. Standard API access runs $1.25 per million input tokens and $4.25 per million output tokens, per Meta's own documentation reflected on OpenRouter's pricing page. A new "Contributor" tier runs $0.10 and $0.20, a 92 percent cut on input and a 95 percent cut on output, in exchange for what Meta's terms describe as permission to train future models on your prompts and completions. The standard tier, by contrast, explicitly excludes your data from training, according to Meta's published terms as summarized by Codersera.

This is not a rounding error dressed up as a discount. On a typical agent turn of 20,000 input and 2,000 output tokens, Codersera's math puts standard pricing at just over three cents and the Contributor tier at under a quarter of a cent, roughly a 14x difference on realistic usage. Meta's AI chief Alexandr Wang confirmed the strategy is intentional, telling press that a "meaningful double digit" percentage of developers have already chosen the Contributor tier and describing the standard rate as steep by comparison, as reported by FelloAI. That is not an apology for the discount. It is a company saying plainly that it would rather have your training data than your dollars, and pricing accordingly.

Here is the steelman, and it is a real one. Trading data for a discount is not a new or shady idea. Search engines, social platforms, and most of the free software you already use run on exactly this exchange, and Meta is at least being explicit about the terms instead of burying them in a privacy policy nobody reads. For a solo developer, a student, or a startup with no proprietary code worth protecting, this is close to free access to a frontier-adjacent coding model, and turning that down on principle is leaving real money on the table for no benefit to anyone. Price discrimination that lets the cash-poor pay in a currency they have plenty of, their own unremarkable side-project code, while the well-funded keep paying cash for privacy, is a more efficient market outcome than one flat price for everyone.

The part that steelman does not survive is what happens once you are not the only owner of the code going through the API. Codersera's review of Meta's documentation found no disclosed retention period, no stated policy on human review of submitted prompts, and no mention of a deletion right once training has happened, adding the plain fact that once material is absorbed into model weights there is no procedure to pull it back out. FelloAI's advice on this point is the correct one: if you work under an employment contract or a client NDA, read those terms before you export an API key that routes through the Contributor tier, because a habit built on a personal project can quietly follow you onto a client's codebase. The discount is real. So is the fact that once you have agreed to it, you cannot un-train what the model already learned from you.

There is a second story from the same week that makes this look less like an isolated pricing quirk and more like a market splitting into two tiers. OpenAI shipped GPT-6 Astra on September 3 at $10 per million input tokens and $50 per million output, reported by Digital Applied, the most expensive rate any ChatGPT-line model has carried, for a capability jump built chiefly around agentic computer use and the first offensive-cybersecurity risk rating OpenAI has ever called its highest tier. That is roughly 2.5 times what OpenAI charges for GPT-5.6 Sol on both input and output, the price I wrote about three weeks ago when Sol itself was the one getting cheaper. Frontier capability is getting more expensive at the top of the market in the same week that commodity coding work got radically cheaper at the bottom, provided you are willing to pay in something other than money. That is not a contradiction. It is two different products responding to two different kinds of demand, and it is exactly what a functioning price mechanism is supposed to do when nobody is fixing the number from the outside.

The practical takeaway for a business evaluating either end of this split: price per token was never the whole cost, and this week made that unusually literal. Before anyone on your team opts into a training tier to save 90 percent on a coding bill, someone needs to answer whether the code running through it is actually yours to give away. That is a five-minute conversation that is much cheaper than finding out the answer after the fact. If you want a second set of eyes on which model tier your team should actually be running work through, that conversation is free.