Open labs need a contributor endpoint

tldr: Meta sells a "contributor" endpoint – roughly 5% of the normal API price in exchange for storing, sharing, and training on your prompts. That’s the honest move, and every open-weight lab should copy it: a direct endpoint instead of an account toggle, discounted capacity instead of hidden training, so your public work funds the labs most starved for agentic data.

Meta just pulled the honest move

Meta did a boss move: it hosts a separate contributor endpoint and sells it cheap. You get two options. Pay next to nothing – call it 5% of the list price – knowing your prompts will be stored, shared, and trained on. Or take the regular paid API and keep your data.

And the strange part: this makes Meta come out as one of the most honest companies in the field. Meta! Of all labs! An honest company, haha.

But don’t underestimate how good this deal is. It’s a win for the customer, a win for the company, and a third win for the field itself – because every API call is a vote on who gets to compete.

That’s why I keep saying that open labs – the ones shipping open-weight models – desperately need to host contributor endpoints of their own. They are the ones most starved for quality agentic traces.

Those traces start in real workflows – genuine error recovery, environment feedback, long chains of tool use – and the standard training diet never reaches them in full.

Not a stealth model buried in OpenRouter, not a random discount promo, but a direct endpoint that behaves like the normal one. You pay less in exchange for your data.

Look at what we have now.

Closed labs charge you full API price and still trade on your data. On a subscription, there is a checkbox informing you that your data will be traded – and you can’t turn it off. Even the plain web interface trains on your data. On the free tier that’s fair: you pay nothing. On a subscription you could argue the fee subsidizes it – but you still can’t switch it off, and it’s account-wide. Not a choice you make consciously each time you run a workflow.

Then the open labs, mostly the Chinese ones: they claim they don’t train on your data, and maybe they don’t, but you can’t check it and you can’t prove it. Everyone knows the possibility is sitting right there the moment you use their native hosted endpoint.

Meanwhile, with open models you typically route through hosting providers that store nothing at all – which is safe, but silent. No signal, no labeled door saying "use me to support this lab."

An openly advertised contributor endpoint is exactly that signal. It tells everyone: this one is specifically for contributing – use it for that. You stop relying on the assumption that a host "probably doesn’t store anything, can’t anyway." One specific endpoint, openly labeled, hands you control over your own risk level. And it becomes the main choice you have: when a lab releases an open model, among all the clouds that merely host it, you can prioritize the lab’s own endpoint precisely because you want to help them.

Why an endpoint, not a toggle

It has to be an endpoint, not a toggle in your account settings asking whether you consent to training. The decision you actually make is per-task: do I route this particular workflow through the cheap endpoint or not.
A toggle collapses that into one account-wide setting you flip once and forget. An endpoint keeps the choice where it belongs – on every run.
A real option – not as :free in OpenRouter thats unsuitable for any long work – one you pay for (even if little) – but without cuts like rate | thinking level | output.

How I route my work

For the last month I used Meta’s Muse Spark exclusively for all of my public projects. If the stuff is going to GitHub anyway, why would I care if my entire history is dumped open? I don’t care at all.

And for many tasks – notifications, scraping, searching – we already operate on the public internet. There is nothing to leak that isn’t already out there.

My commercial work project, meanwhile, stays on providers that verifiably do not store or process my data at all.

The Pirate Bay litmus test

Plenty of things in your business are technically secret yet useless to anyone. So you need a quick litmus test: imagine that every prompt and every result went straight to The Pirate Bay. Would you care, or not?

And even without real secrets, traces can carry identifying crumbs – say, from a personal memory engine – and who knows how those could be connected or used against you.

Yes, this poses a couple of problems you need to address. It does require a little thought beforehand.

Three disciplines before you switch

First, sort by destination. If it’s public anyway, you don’t care: open-source projects, public post writing, scraping, processing of already-public information. Nothing to protect there.
Second, use a provider that cannot connect your prompts to your account. I route through OpenCodeGo over OpenRouter, and the endpoint operator cannot see which account fired the prompt. It can’t be traced back to you, so it all ends up in one big anonymous pile.
Third, the discipline of not dumping your keys into an open chat. Actually, that’s the most important one of all.

Four wins

You get cheaper capacity. Not free, but much cheaper – and when something gets cheaper and accessible, the Jevons paradox strikes and you simply use more of it. I did: I processed more of my old unpublished projects exactly because of that. It got so cheap I stopped checking prices altogether.
You directly cast your vote. Imagine everyone adopted contributor endpoints – you would genuinely become a contributor, supporting the lab with something of value. You vote. You choose. I like this company, therefore I use their endpoint. Even if the model behind it might be slightly worse than the alternative, or they’re roughly the same – you still pick the one you like most.
Because the cheap capacity is less private by design, public-facing projects get encouraged and prioritized. Access to such capacity incentivizes you to build more in public, to build on the assumption that it’s going to be shared anyway.
For open labs exclusively, this is a way to get an edge. They can’t earn enough to pay professionals in the field to prepare datasets, and they don’t have enough compute. This is their way to get ahead and keep improving anyway.

"It’s just subsidized anyway"

You would complain that the cheap endpoint is pure subsidy. Not entirely. Look at the API pricing of proprietary labs: margins run up to 100% across the board. You can check what hosting an open model actually costs – typically between $1 and $3, sometimes far more when you have something in the trillions of parameters – and that gives you a rough idea of real inference cost, plus maybe a 50% provider margin on top. Private labs charge way more than that.

Meta’s own example: the contributor endpoint costs 20 cents of output, the private one 4 dollars. We don’t know exactly what serving the model costs them, but you can assume the private endpoint carries fat margins that possibly already cover the contributor one. Assume raw inference – rent and electricity – runs around 50 cents. Selling at 20 cents means losing maybe half of it per call, while the private endpoint rebounds the cost many times over. Don’t think of it as an immediate monetary loss for them. It’s an investment.

"You’ll just harvest garbage"

The other complaint: you could host such an endpoint and receive nothing but dirt – low-quality data and all that crap inside. Absolutely valid concern, knowing that nothing truly valuable – private business operations, really organic workflows on quality big projects – would ever come in sight.

And yet we progressively have better tools for raw dataset filtering. We now have a wide array of decision models that make decisions fast and help you filter datasets more efficiently. In general, a company with a culture of public endpoints naturally develops more tools to manage that. In a short time you would grow into a wide zoo of classifiers to separate the good stuff from the bad.

The copyright laundering bonus

Here’s the win nobody prices in. Everything submitted through a contributor endpoint goes in under one presumption: the user already holds the legal rights to operate on that data. That’s the agreement – strictly between the user and the endpoint. Whatever the user supplies, the provider gets the rights to, regardless of what it is.

So nobody can ever find out whether the books in the training pot arrived because the lab scraped them by hand or because users were simply processing them for translation or summarization. All of it melts into the same pot, and nobody knows where anything came from.

That’s the great laundering of the copyright crap – the cleanest fuck-off ever delivered to the copycrabs.

Cheap tokens for you, live agentic traces for the lab, legal armor against the copycrabs, and a field where your traffic actually votes. Everyone wins – which is exactly why this shouldn’t stay a Meta exclusive. Somebody build the next one.

We cannot truly be certain, but you may partially attribute Meta’s success to having a public endpoint, or to getting the same data from users anyway. Maybe it contributed enough, maybe it’s a waste, we don’t know.
But I’d still suggest everyone else try to make one and share the findings.