It Was Always an IPO

On June 1, Anthropic filed a confidential S-1 with the SEC at a reported $965B valuation, edging past OpenAI’s $852B for the first time. It could list as soon as this fall, beating its rival to the public markets earlier than anyone expected.

I’ve spent the better part of a year writing about Anthropic’s moves one at a time: the trust erosion, the enterprise tax, the walled-garden crackdown, the Mythos security theater. Each looked, in isolation, like a safety lab making an awkward commercial compromise. The filing is the Rosetta Stone. Read backwards from the S-1, they stop being compromises and start being a sequence.

The Filing Is a Business, Not a Lab

The numbers in the prospectus are not a research org’s numbers:

• Roughly $45B ARR in May 2026, up from $9B at the end of 2025
• Q2 revenue projected at $10.9B, more than double Q1’s $4.8B
• First operating profit projected this quarter (about $559M)
• Gross margin charted from roughly 50% in 2025 to 77% by 2028
• 80% of revenue from enterprise, with 1,000+ accounts spending over $1M a year - a number that doubled in two months

That last line is the whole story. Anthropic is not a consumer company that happens to sell to business. It is an enterprise company that happens to have a beloved developer tool out front. Once you hold that frame, the rest of the playbook snaps into focus.

Pillar One: Safety Was the Moat

The hardest move to read in real time was the Mythos drip-feed. A cyber model too dangerous to release, gated behind a velvet rope, announced with maximum gravity. Then OpenAI shipped comparable cyber capability for $20, and a startup reproduced the flagship findings with $0.11-per-million open models. The gating posture didn’t survive a single news cycle.

If the goal were actually containment, that’s a failure. If the goal were brand, it worked perfectly. “Too dangerous to release” is not a safety control. It is a positioning statement: we are the responsible ones, the adults, the lab governments should trust. For an enterprise buyer signing a seven-figure contract, and for a regulator deciding who gets a seat at the table, that brand is worth more than any benchmark. The mystique was the moat, and a moat is a valuation multiple.

Every AI lab said they were different from OpenAI. Now Anthropic files for IPO. Public markets have a very specific definition of success, and it isn’t “how responsible.”

— A developer on X, the morning of the filing

Pillar Two: The Max Plans Were Loss Leaders

The $200 Max plan never made sense on its own. My own ccusage for May came to about $8,000 in retail-equivalent API tokens, against a $200 bill. What that actually costs Anthropic to serve is contested: Martin Alderson argues the real compute is only about 10% of retail (so about $800), pointing to rivals who serve comparable-size models profitably at a tenth of Anthropic’s list price. SemiAnalysis puts Anthropic’s blended inference margin near 70%, which implies closer to 30% (about $2,400). Either end of that range is a loss of several hundred to a couple thousand dollars a month on a single power user. Companies don’t do that by accident.

They do it to buy mindshare. Get every serious developer fluent in Claude Code, build the habit, build the muscle memory, make it the thing they reach for. The loss is a customer-acquisition cost, and the customer being acquired isn’t the developer paying $200. It’s the enterprise that developer works for.

Pillar Three: The Enterprise Ratchet

Here’s where the slide from 50% to 77% gross margin actually comes from, and it’s been happening in public for months:

• Nov 2025: enterprise pricing shifted from a fixed per-seat fee to a low $20 headline seat plus mandatory consumption commitments - lock in the spend, then meter it.
• April 2026: Anthropic ejected bundled tokens from enterprise seat deals, so usage that used to be included now bills separately.
• Layer on the tokenizer change that can charge 35% more for the same prompt, and killing Opus on cheap plans.

None of these were framed as price hikes. All of them expand margin. An S-1 needs a credible path from 50% to 77%, and this is what that path looks like in invoices: capture the developers cheaply, then make the enterprises they sit inside pay by the token, forever.

What This Isn’t

I want to be careful, because the cynical version of this is too easy and partly wrong.

• This is a read, not a confession. Improving margins as you scale is what every software company does. It doesn’t require a five-year scheme, just a company that consistently chose the commercial option at each fork.
• The safety work is real. Anthropic genuinely does interpretability and alignment research that matters. The argument isn’t that safety is fake. It’s that safety and marketing turned out to be the same department.
• Going public doesn’t kill the mission. It subordinates it. A public company can still care about alignment, but it now answers to a market that prices growth and margin, not restraint. When those conflict, the S-1 tells you which one wins.

The Tell

The most useful thing about an IPO filing is that it forces coherence. A private company can tell a hundred stories to a hundred audiences. A prospectus has to pick one, and the one Anthropic picked is a high-margin enterprise software business with a generational brand and a clear monetization ramp.

Everything else - the gated scary models, the subsidized terminal, the quiet price ratchet - reads, from here, as the work it took to be able to write that sentence. The safety lab narrative had a good run. It was always going to end at a filing.

It was always an IPO.