Stripe's $7B OpenRouter Deal, Cursor's GitHub Rival & the Token Broker Economy - The Weekly Diff #9

Stripe Buys OpenRouter for Over $7 Billion

OpenRouter is joining Stripe, and the reported price is north of $7 billion. For a company that raised $113M only a few months ago, that is a very fast trip up the valuation ladder.

If you haven't used it, OpenRouter is a single API that sits in front of dozens of model providers. You point your code at one endpoint, you get one bill, and you get automatic fallback when a provider is down or a model gets pulled. I called that routing layer essential infrastructure back when a government ordered an Anthropic model offline, because the whole point of routing is that your workflow survives any single provider disappearing. Stripe already runs the billing rail for a huge slice of the internet, and OpenRouter already meters tokens across every provider it touches, so this looks like Stripe buying the meter that sits on top of all AI spending.

Plenty of people admired the developer experience and were happy for the team, while others rolled their eyes at a middleman company being bought by a bigger middleman company and asked, fairly, why a routing layer is worth billions in the first place. The launch post's line about powering "the next wave of GDP growth globally" landed as exactly the kind of corporate language you'd expect to see stapled to a deal this size.

Whoever owns the layer that meters and bills AI usage collects a small cut of an enormous and growing number, and that is a very Stripe thing to want.

The Grey Market Reselling AI Credits

"Who are the token brokers" digs into a grey market where AI credits get resold at a discount, including the $2,500 of credits handed out through YC Startup School and unused API balances people would rather turn into cash. Some of it flows through relays that offer ultra cheap tokens, with the obvious catch that your prompts and traces may be getting harvested on the way through.

When tokens are close to a commodity and providers are handing out credits to bootstrap adoption, arbitrage is the natural end state. Free credits meant to pull developers onto a platform become inventory the moment someone realizes they can sell them.

Selling credits you paid for and won't use feels close to legitimate, even if it technically breaks the terms. Flipping free program credits is a different matter, and risking your standing with a network like YC over a few hundred dollars is a questionable trade. Several people also pointed out how easy this would be for a provider to shut down, since the relay IPs are traceable back to the source accounts, and the version of this happening in China dwarfs anything described here.

Cursor Launches Origin, a GitHub Alternative

Cursor shipped Origin, its own code hosting platform, along with an engineering post on handling Git at any scale. The timing was hard to miss, since it landed right after GitHub's August 17 outage had people openly asking for alternatives, and a fresh reminder that leaning on a single provider is its own kind of risk.

AI coding tools started as editors, then they wanted your terminal, and now they want the repository itself. This is the same direction as GitHub folding stacked pull requests into its own product, except from the other side, with the AI-native tool moving down the stack to own where the code lives. It is worth noting some of the people behind Origin came from Graphite, so the stacked-diff and code-review thinking is baked in rather than bolted on.

Calling the product "Origin" collides head on with origin, the default name for your Git remote. Ask an agent to "push to origin" and you now have two plausible meanings, which is a thin line to walk between a clever growth move and setting a trap where a coding agent quietly pushes your code to a provider you didn't choose. People also had practical questions the launch didn't fully answer, like whether you can browse a repository at all without a Cursor account.

An AI Autofix Opened a Door Into Snowflake's Jira

Wiz's red team traced a path into Snowflake's internal Jira back to a single bad change, and what they found was a GitHub Copilot autofix, generated to tidy up a GitHub Actions workflow, had quietly introduced a command injection bug. They discovered that a single quote placed in an issue title broke out of an echo '...' in that workflow and let arbitrary commands run in the pipeline.

And it wasn't that the workflow didn't have a defensive check. The workflow had an if: condition that checked github.event.pull_request.user.login, which looks protective until you notice that on issues events that field is always null, so the check did nothing at all. Even if it had worked, it would only have excluded a single bot account while letting everyone else through, so it never protected anything.

The autofix was cleaning up a workflow that used deprecated Jira actions, the sort of low priority chore that would have sat untouched in a backlog for a year before anyone shipped it. AI made producing and merging that change cheap, and a subtly broken version of it went in. This is one example of the problems with AI-generated code I've written about before, and it rhymes with the supply chain and injection attacks that keep finding new front doors.

How Bluesky Draws Its Logo on Your Screenshots

Bluesky detects when you take a screenshot and stamps its own logo onto the captured image, and the write up walks through how it pulls that off. X and Threads do the same thing, so this is now a pattern.

The one place people agreed touching a screenshot is reasonable is preventing it as a security measure, in a banking or messaging app, and turning a screenshot into free advertising is a different thing entirely.

When you find engineering effort pointed at controlling how your own content spreads rather than at making the product better, that tells you something about how your product is doing.

A 125M Model That Autocompletes Piano, On Your Device

Someone trained a 125M parameter model to autocomplete piano and run it entirely on-device. You play a few notes, and the model continues the phrase, with no cloud call in the loop.

I love this for the same reason I liked Gemma 4 running in a 2 GB memory budget on a Mac. It is a reminder that not everything needs a frontier model behind an API meter. A small model that fits on your machine can be genuinely useful, and in this case fun, without any of the ownership and portability worries that come with hosted AI.

The response was mostly delight, with a wishlist attached. People wanted to play a melody and get proper multi-part accompaniment back, an export to a MIDI file they could edit, and an option to trade raw speed for musical form, since the model can already emit far more notes per second than any human could play.

That is the diff for this week. Of everything here, the one that stuck with me was a tiny model on a laptop making music for no reason other than that it can.