On February 14, Peter Steinberger posted that he was leaving OpenClaw, the agentic harness he'd built on top of Claude, to join OpenAI. Sam Altman welcomed him publicly. OpenClaw was handed to an open-source foundation with OpenAI's continued backing.
Forty-nine days later, at noon Pacific on April 4, Anthropic sent a message to Claude Pro and Max subscribers: your subscription limits no longer cover OpenClaw or any other third-party agent harness.
Forty-nine days. And this wasn't the first shot. Anthropic had already hit OpenClaw with trademark complaints in January, forcing Steinberger to rename the project twice in one week (from Clawdbot to Moltbot to OpenClaw, because Moltbot "never quite rolled off the tongue"). That was when OpenClaw was just an indie tool built by a solo developer. Then Steinberger walked across the street to OpenAI, and the indie tool became the fastest-growing GitHub project in history: 100,000 stars in 48 hours, 247,000 by March.
If you're an engineer, you already see what happened. If you're a VC, zoom out one more level: both Anthropic and OpenAI are racing to IPO in Q4 2026. Anthropic is targeting an October Nasdaq listing at $400-500 billion, with Goldman Sachs and JPMorgan running the book. OpenAI is reportedly right behind them. OpenAI's board is worried that if Anthropic lists first, it absorbs the pent-up retail demand for AI exposure. Every dollar of margin matters when you're filing an S-1. Every strategic move matters when you're competing for the same pool of public market capital.
Cutting off the subscription subsidy for your competitor's open-source agent framework six months before your IPO? That's not an engineering constraint. That's a calculated play.
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
The math was always broken
Here's the thing about flat-rate subscriptions: they work when usage is bounded. A $200-per-month Claude Max plan was designed for a human typing prompts, reading responses, and occasionally switching tabs to argue on Slack. That's a predictable compute budget.
A single OpenClaw instance running autonomously, 24 hours a day, executing long agentic loops across complex tasks, can consume the equivalent of $1,000 to $5,000 in API costs per day. Per day. Under a $200-per-month plan.
That's not a business model. That's a subsidy.
Anthropic's head of Claude Code, Boris Cherny, framed it as an engineering problem: first-party tools like Claude Code maximize "prompt cache hit rates," reusing previously processed context to save compute. Third-party harnesses often bypass these efficiencies, because they weren't built with Anthropic's infrastructure constraints in mind. Every autonomous run is essentially fresh compute, from scratch, with no caching benefit.
I actually believe him on this. The technical explanation is coherent. But here's the thing: coherent technical explanations and strategic timing can both be true simultaneously. Anthropic was hemorrhaging compute on third-party usage AND that third-party usage was now flowing through a tool backed by their biggest rival, six months before both companies file S-1s. The engineering reason is real. The competitive reason is also real. And when both reasons point in the same direction, the decision gets made fast.
Peter Steinberger said it plainly: "Funny how timings match up, first they copy some popular features into their closed harness, then they lock out open source."
Not a coincidence. A competitive correction dressed up as an infrastructure one.
Platform companies always take back the margin they left on the table
Yep. Always.
Every engineer who has built production systems on top of someone else's infrastructure has a story like this. AWS announces a new managed service that competes directly with your biggest customer's product (Remember the Elasticsearch and AWS OpenSearch debacle?). Stripe adjusts pricing for exactly the transaction type your entire revenue model depends on (get read for token-based pricing). Apple ships a native feature and reclassifies your app category as a setting. The platform doesn't owe you the margin it gave you during its own growth phase, and eventually it comes for that margin.
What OpenClaw ran into is what happens when a platform company decides that subsidizing third-party usage was fine during growth mode but isn't fine anymore, especially when the third party just became an arm (Claw?) of your biggest competitor and you're both about to go public. (In this case: Claude Code, Anthropic's first-party agent harness, launched into direct competition with exactly the use case OpenClaw served. And OpenClaw's creator now works for the other team.)
The question for any startup building on top of an AI lab's infrastructure is not "does this work today." It's "does this survive the infrastructure correction." And the infrastructure correction is not a risk, it's a certainty. The only variable is timing.
Every SWE who has ever migrated a codebase off a deprecated API, or scrambled when a cloud vendor changed a pricing tier, knows this in their bones. The job of a good SWE-background investor is to hold onto that knowledge when someone pitches you a compelling demo that runs on someone else's foundation.
Thousands of users just got a 50x bill. What does your portfolio look like?
The reaction was swift and mostly correct. David Heinemeier Hansson, who has spent years fighting platform dependency in his own battles with Apple's App Store, called the move "very customer hostile." On Hacker News, affected developers calculated that their costs could jump from $200 a month to as much as $10,000 a month to maintain the same usage.
Fifty times.
Anthropic is offering a one-time credit equal to the subscriber's monthly plan, redeemable through April 17. Users who pre-purchase bundles of extra usage can get discounts of up to 30%. Generous, given the circumstances. Not generous enough to offset a 5,000% cost increase.
If you are a VC with AI-native companies in your portfolio, and any of those companies have built their product economics around flat-rate AI subscription pricing rather than proper API billing, you need to have this conversation this week. Not because Anthropic is uniquely villainous (they're not, this is how platforms work) but because every AI lab with a subscription plan and a competing first-party product is now watching Anthropic's move closely, and the infrastructure correction is coming for all of them.
Here's the diligence question I'm adding to my checklist: "If this vendor converted every flat-rate subscription to API pricing tomorrow, does the unit economics still work?"
If the answer is no, you're holding a company whose margins are currently subsidized by someone else's pricing decision. That's not a moat. That's a lease on someone else's moat, and the landlord just raised rent.
The Q1 2026 venture numbers are genuinely astonishing: $300 billion in three months, 81% flowing to AI companies. That capital is going somewhere, and a lot of it is going into startups that are, right now, building on infrastructure they don't own, at prices that were never meant to last. I find the opportunity exciting. I also find the infrastructure dependency risk underpriced in most of the deals I've seen.
Where this leaves me
Honestly? A little chastened.
I don't think Anthropic is wrong to enforce pricing that reflects actual compute usage. A subscription that costs $200 a month and consumes $5,000 a day in resources was a loophole, not a feature. The pricing correction is overdue. But let's be real: the timing wasn't driven by engineering. It was driven by the IPO clock, by Steinberger's defection, and by the fact that subsidizing your rival's ecosystem is a terrible look in an S-1.
And that's the part that should make VCs nervous. Platform pricing changes aren't just about compute economics. They're about competitive strategy, about IPO positioning, about decisions made in boardrooms that have nothing to do with your portfolio company's roadmap. The startup built on top of that platform doesn't get a vote.
I've talked to founders in the past six months where their unit economics assumed flat-rate compute access or subscription pricing they didn't fully control. I didn't always push hard enough on that assumption. The demo was good, the traction was real, the team was sharp, and I filed "compute pricing risk" under "things we'll manage post-close."
OpenClaw is a reminder that "post-close" has a specific date on Anthropic's calendar. And that date isn't set by engineers. It's set by bankers.
The engineers in the room will understand this immediately. The job, my job, is to make sure everyone else in the room understands it too before the term sheet is signed.
SWEdonym