Meta released Muse Spark 1.3 on September 2, its agentic coding model line that the company frames as a step toward what it calls personal superintelligence. The announcement on Meta’s research blog, posted in the early hours of Thursday Beijing time and picked up by Hacker News at 440 points and 297 comments, says the new version improves performance across agentic and coding tasks, drawing on what the team learned from months of Muse Code and Meta Model API adoption. The pitch is about staying power: longer-horizon work, juggling multiple workflows inside a single long thread, generating its own context across messy sources, asking clarifying questions when a prompt is ambiguous, confirming before consequential actions, and following long-form instructions without drifting. A max reasoning mode is still in safety testing, with the rest rolling out today in Muse Code and the Meta Model API.
The thread’s first benchmark was not Meta’s, of course. Simon Willison ran the community’s running canary test within minutes - asking the model to generate an SVG of a pelican riding a bicycle - through his llm tool. The result: 4.2266 cents and 38 seconds, and Willison noted that version 1.2 had animated the pelican without being asked, where 1.3 waited for instructions, which reads as the promised instruction-following improvement. From there the commenters split into two long-running arguments. The first is the pricing model, which is the actual news. The API lists Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens when your data stays private, and $0.10 and $0.20 on the contributor endpoint, where Meta may use your chats and tasks to improve its products. One commenter counted it as more than ten times cheaper than the standard tier; another called it the best price-to-intelligence ratio on the market, beating even legacy DeepSeek V4 Flash pricing; a third said “practically free for contributors at 0.2 usd/mtok, that’s going to be hard to say no to for hobbyists.” The second argument is what the contributor endpoint actually buys Meta. “Privacy is not free,” wrote warkdarrior. “They make it quite clear that they charge more if you don’t want your data used by Meta.” Others were blunter: “they want to train on your chats and tasks and are willing to subsidize for the privilege.” One developer who used the 1.2 contributor tier for real work said it felt like a model that knew its weaknesses, which is exactly the capability Meta says it trained 1.3 to have.
The benchmark claims drew the usual chorus. Commenters reported Muse Spark 1.3 posting a 75.4 on DeepSWE, briefly the best score on the board after Gemini 3.8 Flash had held the top for a few hours the same day, and immediately the thread asked whether either number means anything. “When are we going to stop pretending these benchmarks have any meaning?” wrote notatoad, and gpt5 argued both versions of DeepSWE are likely contaminated at this point, the same fate the pelican test is accused of - Willison’s own benchmark has become so famous that several commenters now suspect labs quietly tune for it, a suspicion that a blog post called “pelicanmaxxing” investigated and largely cleared. The favorite joke of the thread was the developer who claimed a job interview asked him to draw a pelican riding a bicycle: “Aced it, got the job as a senior software engineer.” And underneath the jokes there is the question scotty79 posed directly: if everyone keeps catching up to the frontier within weeks, are we entering a flatter part of the curve?
🎩 Cask’s Take
The story here is not the model. It is that Meta has done something quietly radical with pricing: it took the oldest deal in consumer tech - your data in exchange for a service - and printed a price tag on it. $1.25 per million tokens if you want privacy, $0.10 if you do not. That is not a discount, that is a posted rate for being training data, and it is the most honest pricing page in the industry right now. Every other lab trains on user data inside opaque terms of service; Meta just made the trade legible and found that the market accepts it instantly. The commenters who shrugged - the hobbyist who said his personal projects are all AI-generated anyway and it would be hilarious if Meta trained on AI-generated code, the developer who said the useful data is not the code but the session transcript, your corrections and refactors - they are not naive. They have just correctly priced their own judgment at somewhere below twenty cents.
That is the part worth sitting with. What the contributor endpoint buys Meta is not your code, which is mostly public anyway, but your steering: the moments where you tell a model a different approach would be better, where you pull it back from a wrong direction, where you accept or reject an edit. That is distilled human judgment, collected at scale, for fractions of a cent per interaction, and then fed back into a model that Meta sells at a price no one else can match because its raw material is subsidized by its own customers. The flywheel is no longer a metaphor you have to squint to see - it is a line item. The only open question is the one the security-minded commenters raised: what happens when the thing people feed the cheap model is not judgment but secrets, and whether a model trained on a million messy sessions starts leaking the mess. For now, the market has voted. At $0.10, the answer to “would you like to be the product” appears to be yes, as long as the product is a good deal.