The 2% Experiment That 98% of Customers Saw | Edition 305

On April 21, 2026, Anthropic changed one line on its public pricing page. Claude Code — the coding tool bundled into the $20-a-month Pro plan — stopped being listed as included. Read the page cold and it looked like the cost of getting Claude Code had just gone from $20 a month to $100, because the next tier up was now the cheapest place it appeared.

It was not a price increase. It was an experiment running on roughly 2% of new signups.

The other 98% found out anyway — because the page they read is the same page everybody reads.

The whole thing was reverted inside a day, which is exactly why it is worth studying rather than piling on. This is a company that is very good at this, making a small, recoverable, entirely ordinary mistake that almost every product team is one sprint away from making.

What actually happened

The sequence, as far as the public record actually supports it. Times are UTC, with the US Pacific equivalent where it helps — because most retellings of this episode got the order wrong, and the order is the story:

• 21 April, between 14:10 and 19:25 UTC (7:10am–12:25pm Pacific). Internet Archive captures bracket the change: at 14:10 the Pro card on claude.com/pricing still reads Includes Claude Code; by 19:25 that line is gone.
• The docs move alongside it. A support article previously titled Using Claude Code with your Pro or Max plan becomes Using Claude Code with your Max plan, captured under the new title at 18:36 UTC on 21 April. The earlier wording survives in an archive snapshot from 16 April.
• 21 April, afternoon and evening Pacific. Developers notice. X, Reddit and Hacker News fill with screenshots. Nothing has been announced, so every account of the change is somebody's screenshot.
• 21 April, 22:55 UTC (3:55pm Pacific). Amol Avasare, Anthropic's head of growth, posts on X: “For clarity, we're running a small test on ~2% of new prosumer signups. Existing Pro and Max subscribers aren't affected.” This lands before the press coverage, not after it.
• 22 April, 01:16 UTC — the evening of 21 April in the US. Hours later, The Register publishes Anthropic tests how devs react to yanking Claude Code from Pro plan.
• 22 April, 02:51 UTC (7:51pm Pacific on the 21st). Avasare again: “Getting lots of questions on why the landing page / docs were updated if only 2% of new signups were affected. This was understandably confusing for the 98% of folks not part of the experiment, and we've reverted both the landing page and docs changes.” Twenty minutes later, replying to Simon Willison, he adds that it “was a mistake that the logged-out landing page and docs were updated for this test.”
• Reverted between 00:12 and 05:10 UTC on 22 April. The archive still shows the line missing at 00:12 and restored by 05:10 — 10:10pm Pacific on 21 April. The support article is back under its original title by 06:39 UTC.

Read in Pacific time, where Anthropic works, the entire episode — change, discovery, coverage, correction, revert — happened on 21 April.

As of today, claude.com/pricing lists Claude Code as included in Pro: $20 a month billed monthly, or $17 a month on the annual plan, which is charged as $200 up front. Max still starts at $100 a month. The revert held.

The experiment wasn't the mistake. The surface was.

Running a price test on a slice of new signups is completely ordinary. Every subscription business does it, and doing it on new signups only — leaving existing subscribers untouched — is the considerate version. On the merits, that part was fine.

The mistake was in the plumbing. The experiment was implemented by editing two surfaces that have no idea who is looking at them:

• The logged-out landing page — the one every prospective customer, journalist and competitor sees.
• The public support documentation — which is also what search engines index, and what an AI assistant reads when somebody asks it what a plan includes.

A cohort-targeted experiment needs a cohort-aware surface. A logged-out page is the definitional opposite of that: it cannot know whether you are in the 2%, because it does not know who you are.

So the reach of the change was never 2%. It was everyone who visited — and then, because screenshots travel, everyone who read about the people who visited.

The rule worth stealing: the blast radius of a change is set by the widest surface you edit, not by the size of the group you meant to target. When those two numbers disagree, the surface wins, every time.

What Anthropic got right, and what it still cost

Three things went right, and they are the reason this reads as a stumble rather than a scandal.

• Existing customers were said to be protected, early and clearly. Within hours, Anthropic's head of growth stated publicly that existing Pro and Max subscribers were not affected. Taken at its word, nobody's bill changed and nobody lost a tool they were already paying for — though that assurance is the only one on the record, and the company published nothing to verify it.
• The revert was fast and complete. Same day, both surfaces, no negotiation. Plenty of companies would have spent that day drafting a defence of the change instead.
• The correction was in plain language. “we've reverted both the landing page and docs changes” is a sentence a human wrote, and calling the whole thing “a mistake” outright is rarer still. Nobody reaffirmed a commitment to transparency.

What it cost is subtler, and it is the part worth taking seriously if you run a product.

For several hours on 21 April, the most authoritative page describing what Anthropic sells disagreed with what Anthropic actually sells — and disagreed in the direction of a fivefold increase. Every screenshot taken in that window still exists. Search engines and AI assistants both formed their answer from whatever the docs said at the moment they looked. A page you revert is not a page you un-publish.

And the statement itself, however honest, came from an executive's personal account on X. Every quote in this edition traces back there. There was no blog post, no changelog, no support note — nothing the company controls, and nothing anyone outside it can check the 2% figure against. Simon Willison said as much at the time: “I don't buy the ‘~2% of new prosumer signups’ thing, since everyone I've talked to is seeing the new pricing grid.” That channel was fast and it worked — but it means the durable record of what happened lives on a social platform rather than anywhere the company controls.

Two numbers to be careful with

This story picked up two figures on its way around the internet, and neither survives a check. We are flagging them because you will run into both.

“Nine hours.” Or twelve. Several write-ups put a precise duration on the episode. No primary source supports any specific number. The archive captures only bracket it: the line was on the pricing page at 14:10 UTC on 21 April and gone by 19:25; still gone at 00:12 UTC on 22 April and restored by 05:10. That puts the change live for somewhere between roughly five and fifteen hours — a range wide enough that any precise figure is invented. A number that specific, with nothing behind it, is how a tidy story beats a true one.

“A $61.5 billion company.” That was Anthropic's valuation in March 2025, and it has been copied forward ever since. By the time of this incident the company had closed a $30 billion Series G at a $380 billion post-money valuation, in February 2026. Five weeks after the pricing page was reverted, it raised $65 billion more at $965 billion post-money, on run-rate revenue it said had crossed $47 billion.

Which is the more interesting version of the story anyway. A company operating at that scale, moving that fast, still shipped a two-line copy change to a public page without anyone asking who else reads it.

What to do this week

If you ship anything customers pay for, this is a fifteen-minute exercise, not a project.

• List your logged-out surfaces. Pricing page, docs, help centre, plan comparison table, FAQ, app store listing. These are the pages that cannot personalise, and therefore cannot be tested quietly. Write the list down once — it is shorter than you think, and right now nobody on your team has it.
• Add one line to your experiment checklist: which surfaces does this change edit, and who can see each one? That single question would have caught this.
• Decide the announcement before the change, not after. The reason this became a story is that the first public account of it was a stranger's screenshot. If a change is defensible, saying it first costs nothing. If saying it first feels bad, that is information about the change.
• Review docs and pricing together. Here they at least moved as a pair. The more common failure is worse: they disagree for months and nobody notices.

The uncomfortable version of the lesson is that this was not a failure of judgement or of ethics. Someone made a reasonable decision to test a price, and the change reached a hundred percent of an audience it was designed for two percent of. That gap is a plumbing problem — and plumbing problems never announce themselves. They just quietly set the blast radius while everyone is discussing the strategy.

Sources: Simon Willison, “Is Claude Code going to cost $100/month? Probably not” (22 Apr 2026) · The Register, “Anthropic tests how devs react to yanking Claude Code from Pro plan” · Ed Zitron, Where's Your Ed At, for the archived support-doc wording · claude.com/pricing, checked 6 Aug 2026 · Anthropic, Series H announcement (28 May 2026) and its Series G announcement (12 Feb 2026).