Anthropic’s Sonnet 5.5 announcement page sits at the top of Hacker News this morning — 504 points, 340 comments, and a thread I would bet real money is mostly an argument about a benchmark table. Everyone is asking whether the model is better. Almost nobody is asking why it has a decimal point in it.

That decimal is the story.

The Decimal Is the Deliverable

A version number is not a measurement of intelligence. It is a promise about paperwork.

A “6.0” tells a customer that something structural changed — the kind of change that invalidates the regression suite, reopens the security questionnaire, resets the vendor risk file, and forces the platform team to re-run evaluations it spent a quarter building. Anthropic’s “5.5” tells the same customer the opposite: same family, same interface, same data-handling terms, better weights. You can adopt that inside a change window. You cannot adopt a 6.0 inside a change window. You adopt a 6.0 as a project — with a kickoff, a rollback plan, and somebody’s name on it.

Which means the release cadence now being set by the labs is not set by capability curves. It’s set by how often a large organization can be persuaded to re-certify a vendor. That is a much slower clock, and it is the one that governs.

Your Risk Committee Is the Real Regulator

The standard right-of-center read on any AI announcement is that the brakes are in Brussels or Washington, and if we could just get the regulators out of the way, adoption would accelerate. I’ve made versions of that argument myself. It’s mostly wrong, and a point release is the proof.

No agency requires a Fortune 500 to re-run its prompt library. No statute mandates the fourteen-page model-risk questionnaire, or the two-week change advisory board, or the fact that the person who signs off on production deployments is a mid-level manager whose actual incentive is to not be the reason something broke. Deregulation does not shorten that meeting. Nothing in the federal register touches it. The binding constraint on AI diffusion in America as of September 2026 is a private-sector risk function that answers to its own lawyers — and it is stricter than any statute, because it offers no appeals process and no due process, just a signature somebody has to give.

The 340 comments on Hacker News are the same phenomenon wearing a hoodie. The industry has quietly replaced the press release with the argument about the press release. Post a model, collect your points, watch the thread. It’s cheaper than a launch event and it produces the same artifact for the procurement file: evidence of market attention.

The Moat Is the Prompt Library You Already Wrote

Here is what a benchmark table cannot capture. The reason an enterprise stays with the model it has is not that the model is best. It’s that the switching cost is enormous and the benefit is a number on a chart nobody in the room fully trusts.

The people doing that math don’t write blog posts. “We ran the swap on our top forty workflows,” a model-risk reviewer at a mid-size bank wrote to me in a Slack message this week, “and eleven of them regressed in ways our evals caught and four regressed in ways they didn’t. That’s not a migration, that’s a year.” That is the actual competitive terrain — not the leaderboard, but the eval suite the customer wrote for the last model and now cannot afford to discard.

This is also the most under-rated answer to the bubble question. A business whose customers renew because leaving is a twelve-month engineering program is not a speculative business. It’s a switching-cost business — duller, and much sturdier. Anthropic and its competitors don’t have to be brilliant every quarter. They have to be stable every quarter and interesting every eighteen months.

What Would Actually Loosen the Bottleneck

If you genuinely wanted to speed adoption — and a functioning policy apparatus would — you wouldn’t preempt state rules or subsidize data centers. You’d build a portable way to certify that model B behaves sufficiently like model A on a given workload, so the eval suite transfers. Every dollar of friction in this market is a dollar of re-validation.

Nobody is building it. The labs don’t want their customers to have a cheaper exit. The customers’ risk functions don’t want to surrender the jurisdiction that friction grants them. And the political class is busy arguing about whether the technology is real.

So we get another point release, another 504 points, another 340 comments. The model gets a little better. The paperwork, which is what actually governs, stays exactly where it was.