I cancelled Claude Max. Not because of the model

I cancelled Claude Max today.

This is not a model ranking. Claude is still very good at writing code. It’s that Anthropic’s latest moves —especially the watermark on text and files— were the last straw. I’m sticking with Grok, for now.

On Tuesday I already wrote why that watermark looks like a dumb idea: it biases the generator, falls over after a paraphrase, punishes whoever complies, and creates false security. This is the personal follow-up. I pay a subscription so the code and the copy I put my name on are mine. Not so they leave with a statistical filigree that rides along in copy-paste.

What it is (and isn’t) #

Anthropic have explained it themselves. On the new models, generated text carries an invisible mark in the word choices themselves —not a logo, not a footer. On files, signed metadata. They apply it worldwide, not only in the EU, because Article 50 of the AI Act is already in force and they don’t have a clean way to scope it by region.

It does not say “Anthropic is the author”. It says, roughly: this went through Claude. Fine for an internal email. For a Moodle plugin, a post on this blog, or an app I ship under my name, it doesn’t sit right.

Code, they admit, is an awkward fit for the scheme (low entropy: the next token is almost determined). That’s exactly why I use it: so it helps me write the right next token, not a token that also doubles as a stamp.

This is not a hate post #

I paid for Claude Max for a long time. For months it was the tool that fitted programming best. I’m not leaving because “a shinier model dropped”. I’m leaving because the product I buy changed the contract: it now includes a marking layer I didn’t ask for and can’t turn off.

I get the AI Act. I get that they want to comply. I don’t get choosing the path that reaches into the sampling of everything that comes out —including what I review, rewrite and sign— and selling that as transparency for the reader.

Transparency, in my case, was already there: I put my name on it. If a piece was written straight by a model, say so. If it’s a diff I went through line by line, it’s mine.

Why Grok, and “for now” #

In July I wrote about Grok with caveats: an aggressive coding model, cheap, and a Grok Build privacy scare I didn’t swallow. I didn’t come in as a fan. I came in by using it.

What kept me isn’t a benchmark. It’s the weekday.

It stays in the repo. One thread, several projects: LiftApp, Cruce, a post, a deploy. I don’t hop from chat to IDE to another tab to make it compile. It opens Xcode, reads git, writes the commit, starts Hugo. For how I work, that outweighs a point on a leaderboard.

It programmes the way programming actually happens. Tests, pbxproj, HealthKit, a Hugo hook, a notarised zip. It gets things wrong, like all of them. It doesn’t stop at the nice comment: it gets to the diff.

It doesn’t mark the text. For now there’s no filigree on the code or on this post. That’s the contrast, not a slogan.

It’s been useful enough to build on. GrokUsageBar exists because I use it enough to want the weekly percentage in the menu bar. If it didn’t fit, there would be no app.

And the model helps. 4.6 shipped on 12 August (announcement), aimed at long-running agents and staying inside a codebase. That’s what I actually notice: it doesn’t tap out on the third tool. Musk says 4.7 is three or four weeks out and will be significantly better. We’ll see: Elon promises dates the way other people order coffee. If it lands and holds up, good. If not, 4.6 is already enough for the work I’m doing now.

It isn’t a marriage. If Grok (or anyone else) ends up in the same place —a global filigree, no switch, on the code I sign— I’ll do the same. Cancelling isn’t an identity. It’s a line.

Meanwhile: I cancelled Claude Max. I’m staying with Grok. We’ll see.