The Mark and the Meter: A Fortnight of AI Gripes

Some fortnights the complaints scatter across a dozen topics; this one they clustered around two. Across the forums where paying AI users gather to compare notes, two things happened in the same window that left the same taste: the meter moved, and the mark landed. In both cases, something about the product changed without a clear announcement, and the users found out by running into it.

Quotes sourced from: the OpenAI Developer Community and Hacker News. Every quote below was opened at its permalink and copied verbatim; each is listed with its handle, platform and date in the Sources section. We quote experiences, not verdicts — a forum post is one person’s felt reality, and we have framed it as exactly that.

The $200 plan that lasted two days

The loudest note this fortnight was the Codex usage drain. On the OpenAI Developer Community, a Pro subscriber posting as jjjnoronha laid out the damage plainly on 4 August:

“I’m on the $200 Pro plan, and my weekly limit was completely exhausted in 2 days of serious development work. This is simply not viable.”

He was refactoring a Rust backend with two agents running — one coding, one checking code quality — and described leaving Ultra mode on for three or four hours, which alone consumed 30% of his weekly allowance. But the complaint was not that the work was expensive; it was that the plan was sold with language suggesting far more headroom than what the cap delivered. “When someone pays for the highest consumer tier,” he wrote, “the expectation is capacity for sustained, professional work — not being locked out for the rest of the week after two days.”

Others described something stranger. A user posting as kalani reported going to bed with 70% remaining after a Saturday reset, then waking up to find 0%. “Not like I had something running overnight, and yet all usage blocked until this next Saturday,” they wrote on 4 August. “Hard to trust a $200 / month service that is completely unpredictable about whether you’ll even be allowed to use it for most of the week.”

What followed was more striking still. On 5 August, kalani reported opening a second $200-per-month account to work around the drain:

“I opened a second $200 / month account Yesterday and went to bed with 91% usage remaining. NOTHING ran overnight, and my laptop was asleep… This morning, I woke up and this account too shows 0% Usage remaining until Monday, August 10 at 11:40pm. So, paid an extra $200 for a second account to get 9% usage for the week, and now BOTH accounts have 0% remaining.”

Two Pro accounts, $400 a month between them, both showing zero remaining within a day. The same user drew the uncomfortable comparison: “You know you’re in a bad spot when a $30 grok subscription looks generous by comparison and letting me do considerably more than I can do with 2x OpenAI Pro $200 subscriptions.”

The limit that keeps changing shape

The individual drain reports were concerning; the pattern underneath them was worse. On 14 August, a Codex 20x subscriber posting as anil.c1 described the shift in more measured terms:

“Back when we had the 5-hour limits, I was actually able to get more work done, and even when I used Codex very aggressively and heavily, my limits would rarely run out in less than five days. Now, the weekly limit seems to drain almost as quickly as the old 5-hour limit did.”

Their usage had not changed. No MCP, no special tooling, the same kind of workload as before. Yet the limits were consumed faster. In the same thread, a user posting as Sandemann captured the behavioural consequence: “You feel forced to constantly squeeze the maximum out of your tokens because you can never be sure when a reset is coming, or if one is coming at all.”

The underlying complaint was legibility. A commenter posting as mat.eo put it concisely on 6 August: “We have no understanding of what the limits are outside of a percentage.” No absolute numbers, no clear documentation of the weekly cap, no way to predict what a session will cost in terms of the allowance. Another poster, plutavian, went further: “It’s not that usage has increased; rather, the weekly usage limit was reduced after the reset.”

And then there was the billing. A user posting as Yottalionaire1 described, on 7 August, what they labelled evidence of a system bug: “26 auto-reloads in 44 hours totaling $466.82 on a $200 Pro Account.” At peak, they were billed at 4:10 PM, 4:32 PM and 4:50 PM consecutively, while the front-end interface hung on “thinking” loops without producing output. Whether this was a bug or an expected consequence of the billing model, the user could not tell — which is the point.

The pattern is not new; we have described before how these plans’ limits are hard to reason about, and how the flat subscription is quietly becoming a meter. What this fortnight adds is the sense that the meter is not only hard to read but actively changing what it measures.

The mark you did not ask for

The second cluster came from a different direction entirely. When Anthropic announced it would embed invisible watermarks in all Claude output — text and files, worldwide, as part of its EU AI Act compliance — the reaction on Hacker News was immediate and pointed. A thread on 16 August drew over 200 comments, and the concerns were not about the regulation but about the implementation.

A commenter posting as nomel framed the ownership question: “So, you don’t own the generated text, and can’t use it freely then. What if I copy paste a section, or rewrite a section of text to my liking? What if I rewrite some lines of code that contains the mark?”

The worry was sharpest for users who treat Claude as a writing tool, not a writing replacement. If you wrote your own text and asked Claude to proofread it, the watermark — which operates at the token-selection level — would mark the output as AI-processed regardless of who actually composed the words. A commenter posting as stabbles noted the irony: “Claude’s writing was already easy to recognize. The fact that Anthropic complied without complaint makes me wonder if they already watermark their outputs and used the opportunity to create goodwill.”

Another commenter, posting as 4d4m, drew the competitive conclusion: “Reminder: your favorite distilled model does not treat you, the customer, as an adversary and mess with your output… May the free market win.” And bushido added, with a directness the thread rewarded: “This is not meant to be snarky, But almost any writing done by Claude is a perversion of writing. I honestly can’t stand the way Claude writes. This watermark change just makes it scarier.”

What the two have in common

Step back from the individual gripes and a shape emerges. One company changed what a $200 plan delivers; the other changed what Claude’s output contains. Neither sent a memo that users would have noticed before hitting the wall. Both left paying customers in the same position: discovering the terms had shifted by running into the consequences, then working backwards to understand what happened.

The connecting thread is not malpractice — we are not saying either company acted in bad faith. Codex limits genuinely have to flex under load, and the EU AI Act’s transparency obligations are real. The thread that connects the two is legibility:

• The meter moved. Codex usage limits shrank or fluctuated without announcement, and users were left reverse-engineering their own plan’s terms from a percentage that changes what it measures between resets.
• The mark landed. Claude began watermarking all output, including edited text, without user opt-in or opt-out, and the detection tools that would let users check their own text are not yet public.
• The memo never came. In both cases, paying customers found out by running into the wall, not by reading a changelog.

A limit you cannot predict, a watermark you cannot see, and a product that changes underneath you without a changelog all produce the same feeling: that you are not quite the customer you thought you were. We have made this argument before about AI token pricing; it applies with equal force to usage caps and output marks.

If there is a practical note to end on, it is the familiar one: keep your own receipts. Screenshot the usage page before and after a session. Read the terms rather than trusting the vibe. And when a product shifts under your feet without telling you, log the date and the detail — because the pattern is the evidence, and individual posts are how it gets documented. Nobody writes a forum post about a product they do not care about.