Same Card, Two Weeks Later: Four Ratings Moved And The 95% Vanished

On 2 October a card called “Save Time with Claude — Welcome to Claude Maxxing” went up: three stacked tiers, Chat, Cowork and Code, eighteen feature tiles, every tile carrying a star rating out of five. On 16 October the same card went up again. The caption was identical, word for word, down to the punctuation.

The card was not. It has a new design, four changed star ratings, one deleted statistic, and one new line of fine print. We downloaded both images and read them tile by tile. This post is about the differences, because a revision tells you something the original cannot: it tells you what the person who made it decided to change their mind about.

The caption is identical. The card is not.

Here is everything that moved between the two issues. Fourteen of the eighteen tiles, and the star ratings on all three tier headers, are unchanged.

Four ratings moved. What happened in those fourteen days?

A star rating is a claim about how good something is. If the rating changes, something changed. The honest question is what.

So we went to the published release notes for the Claude apps and read everything dated between the two posts. There are two entries, both on 7 October: Claude Haiku 5.5 launched, and Max and Team plans gained monthly API credits. That is the whole list.

Neither touches Scheduled Tasks. Neither touches Plugins. The most recent plugins entry is 25 September — a developer portal for submitting plugins to the Claude directory — and that is a week before the first version of the card, so it cannot explain a change made after it.

That last row is the clean one. Ask User Question is not a feature at all. By the card’s own description it is a sentence you add to a prompt. Nothing can ship that makes it work better or worse. Its rating went from four to three anyway — which tells you the ratings are not tracking the product.

Diagram: a star rating can move for three reasons

When a number on a card changes, there are only so many explanations. Lining them up against the evidence is the whole exercise.

The number that left the card and stayed in the caption

The first card said it inside the Chat tile, in the artwork, where it would be screenshotted and shared: “95% of users stop here and never level up.” The second card says “Many users stop here” instead.

That is the right edit, and it is worth saying why. We went looking for a published figure splitting Claude users into chat-only and everything-else. There is no such figure from Anthropic. What they have published about usage is a study of Claude Code covering roughly 400,000 interactive sessions from about 235,000 people between October 2025 and April 2026 — a study of the people who are using Claude Code, which by construction cannot tell you what share of everyone else is not.

So “95%” was a figure with nowhere to point. Taking it out of the artwork is a genuine improvement.

The complication: the caption still opens with it. “95% of Claude users stop at Chat.” Same caption, both posts. The number was removed from the part that gets saved and re-shared, and kept in the part that gets scrolled past. If the figure was not solid enough for the card, it is not solid enough for the first line either.

The fine print is the best thing on the new card

The 2 October version had no disclaimer at all. The 16 October version has this, in small grey type at the bottom:

“Menu names, features and plans change often. Check Claude’s official site for what is available to you. Star ratings are a rough guide, not a measurement.”

That second sentence does more work than the whole redesign. It tells a reader exactly how much weight the numbers will carry, before they start adding them up. Most cards in this genre present a score and let you assume it was arrived at by scoring something. This one says plainly that it was not.

It also quietly resolves the puzzle in the section above. A rating that is a rough guide does not need a product change to move. It only needs the person holding the pen to feel differently about it on a Thursday than they did a fortnight earlier. That is a legitimate thing for a card to be — as long as it says so, which this one now does.

What the arithmetic says after the revision

When we looked at the first version we added up the tile ratings inside each tier and found that the ladder the card draws does not match the scores the card gives. That result survives the revision, but only just.

Each of the four edits pushes in the same direction: Cowork up by 0.33, Chat down by 0.08, Code untouched. If you were trying to make a two-four-five ladder look like a two-four-five ladder, these are exactly the four edits you would make.

And it still does not get there. Level 2, the rung the caption is urging you to climb to, is still the lowest-scoring tier on the card by its own tile ratings. We worked through why that is in the piece on the first version, and the short answer has not changed: the tiers are not sorted by how good the features are, they are sorted by how much you have to install and own before a tile does anything.

The useful habit here has nothing to do with Claude

Cards like this one get re-posted constantly, and almost nobody compares the versions. It takes about five minutes and it is the cheapest quality check available to anyone who follows a page for information.

1. Save the image, not the screenshot of the feed. You need the artwork at full resolution to read small type and half stars.
2. Diff the numbers first. Numbers are the part that is supposed to be stable. If they moved without the product moving, you have learned what kind of document you are reading.
3. Diff the disclaimers. A disclaimer that appears in version two is usually a direct answer to something version one got pushback on.
4. Check the caption against the card. They are written at different moments and they drift. Here the card dropped a statistic the caption kept.

Four honest caveats

1. Release notes are not a complete record. Software changes that nobody writes up. Our finding is that there is no published change in the window for these four tiles — which is weaker than “nothing changed”, and we are not claiming the stronger thing. This is the same discipline we wrote about in “not reported” is not “not there”.
2. Half stars are hard to read. We zoomed to full resolution on the Connectors tile to confirm the half star on the 16 October card and the whole star on the 2 October one. The other three changes are whole stars and unambiguous.
3. We do not know the maker’s reasoning. Everything above is read off two images. A rating may have moved because of hands-on use we cannot see, which is a perfectly good reason and one the new fine print explicitly makes room for.
4. This is not a complaint about the card. Version two is better than version one: it removed an unsourced statistic and added a line telling you how much the numbers are worth. Those are the two edits a careful person would make. The point of this piece is that the edits are visible, and visible edits are information.

The one-line version

The same card ran twice in fourteen days with an identical caption; four star ratings moved with no published product change behind any of them, the one statistic was deleted from the artwork but left in the caption, and the new fine print now says out loud that the stars were never a measurement.

Which makes the most honest sentence on the second card the one in the smallest type. If you are going to take one thing from a card that rates eighteen things out of five, take that.