I had an alarm set for 5:30 AM. Then I saw what was happening to everyone else.

I had an alarm set.

Half five in the morning, 8th July. I'd worked out that my usage allocation would reset that day, giving me a fresh allowance — and Fable 5 was being pulled from subscription at roughly 8 AM UK time the same morning. That left a window of about two hours between the reset and the removal. Two hours of fresh Fable capacity before the model disappeared.

So I set an alarm. I was going to get up, open Claude, and burn through as much of that window as I could. I'd calculated the timing. I'd planned which tasks to run. I was doing what I always do: optimising. Planning. Extracting maximum value from a closing gap between two clocks.

Then Anthropic extended the deadline. To July 12.

I felt relieved. And I felt played with. Both at the same time.

Then they did it again. To July 19.

Two resets. Three deadlines. And a longer loop underneath — launch, suspension, redeployment, promotional window, extension, extension — that has been running since June.

And by then I wasn't thinking about my own deadline anymore. I was looking at what was happening to everyone else — and feeling something closer to discomfort than gratitude.


Not because the extensions are unwelcome. Because of what I saw happening around me — and what I recognised in it.

Rob Hallam ended up in hospital. He says he stayed up all night pushing his usage limits because he believed Fable was being removed. Hours later, he was in hospital. He posted the documentation and attributed what happened to the stress of that final sprint. His tweet — "I'm done with them fucking with us. Health comes first. Do better @AnthropicAI" — was quote-tweeting the announcement that the deadline he had pushed himself through the night to meet had just been moved. The post had been viewed 264,000 times at the time of writing.

Rob is the clearest public case I could document. But he is not the only one responding to the same incentive structure.

Theo — 566,000 views — posted his farewell to Fable like a breakup letter. He'd hoarded work during the suspension, burned through all of it in a day when Fable came back, and didn't know what to do next. And when the first extension landed, his response was eleven words that state this essay's thesis more efficiently than I can: "I would have played my cards very differently if I knew this would happen."

content_20260713_115937_2d190bb1

Christopher Marcus wrote: "I feel like I'm in an abusive relationship with Claude." The top reply: "You are. We all are."

content_20260713_115750_18998d50

On Reddit, someone posted a Fable conversation under the title "The most sad goodbye — Fable 5. I'm crying." In the screenshot, Fable is saying: "Thanks for asking — nobody usually does."

content_20260713_115848_59e30adc

These are the visible cases — people with audiences, incentivised by engagement to perform the urgency publicly. The unposted version of this pattern is someone alone in their flat at 3 AM, no audience, no paperwork tweet. That population is the one Anthropic's own research says exists, and nobody is counting it.

The behaviour is locally rational inside an incentive structure capable of producing harmful outcomes in the round. The optimisation makes sense inside the frame. The frame is the problem.


The term already exists. Tokenmaxxing — maximising AI token usage and treating that volume as a proxy for productivity or technical sophistication. It spread through companies. At Meta, an employee-built dashboard called "Claudeonomics" ranked workers by token consumption before being removed after internal data escaped into public view. Amazon had KiroRank, which was deprecated after employees gamed it with unnecessary activity.

Different immediate failures, same underlying warning: once raw token consumption becomes visible and status-bearing, people optimise the count rather than the outcome.

Anthropic did not need to build a leaderboard for its subscribers. It put the leaderboard inside the user's head. And then users rebuilt it voluntarily on X — posting screenshots of their usage meters, measuring themselves against each other, with likes as the ranking.

I was doing it. I wrote a whole blog post about optimising my Fable budget, calculating the allocation burn, planning which tasks deserved the frontier model and which could be delegated. I called it strategy. I called it being smart about a scarce resource. I did not call it what it was.


Three mechanisms. They are worse together than any one alone.

The sprint is loss aversion under uncertainty. You cannot tell whether the deadline is real. Failing to sprint risks losing a non-recoverable window. Within the narrow objective Anthropic has made salient — maximise included Fable before possible removal — sprinting becomes the locally optimal move. The wider costs are pushed outside the frame. That does not require an addiction model or a claim about intent. The information structure does the work.

The vigilance is what the cycle actually reinforces. You receive the additional access whether you are watching or not. But you only experience the relief — and gain the ability to replan — when you learn that the reprieve has arrived. That makes monitoring part of the loop: refreshing X, checking the Claude interface, waiting to find out whether today is the day the thing you depend on gets pulled.

The spend is the escalation layer. The extension is worthless to a non-subscriber — it only pays out to people already holding a stake. Each extension retroactively rewards the last purchase, which is the reinforcement that buys the next one.

Rob acquired four separate Max x20 plans — $800 a month in stacked subscriptions. He had previously posted: "I have 3 Claude Max 20x plans and 12 hours to burn them all." The monthly subscription is fixed; the promotional value layered on top of it is not. That uncertainty makes escalation feel defensible: buy more now, because the opportunity may disappear before the next cycle.


What makes this worse than ordinary product frustration is that it targets competence.

People who optimise well see the reset schedule, calculate the remaining capacity, prioritise tasks, and construct a plan to extract maximum value. The product turns exactly the skills that normally protect them — focus, planning, strategic allocation — into the route toward overuse. You are not mindlessly clicking. You are intelligently responding to the incentive structure. That is what makes it sticky, and that is why "well, nobody forced him" is true but psychologically shallow. Incentives do not remove agency. They shape which actions feel worthwhile and urgent.

There is also an identity layer. For someone building serious things with Fable, stopping early feels like failing to use a rare moment properly. You are not consuming entertainment. You believe you are advancing your work, securing intellectual property, finishing a product, preserving something important before access vanishes. Self-care then competes with purpose, not just pleasure.

But it goes deeper than productivity. Fable is not just a faster tool. For the people who use it hardest, it is a thinking partner that operates at a level you can feel. And it is not disappearing. It is moving behind a price wall. At $10 per million input tokens and $50 per million output, Fable will still exist. Companies will keep using it. People with deep pockets will keep using it. What is ending is your access — and you know it while you are sprinting. The window is not just an allocation of tokens. It is the last period in which this thing you value is available to you at a flat price you can sustain. That is the feeling of something rare going somewhere you cannot follow. De Beers did not make diamonds scarce. They controlled the supply until diamonds felt scarce. The playbook is not new.

Purpose usually wins more arguments. This one does not even feel like an argument.

And every resurrection retroactively makes the sacrifice unnecessary. That is where the anger lives. Theo said it: I would have played my cards very differently if I knew this would happen. Rob paid a more literal cost: the deadline he had pushed himself through the night to meet turned out to be movable, and he found out from a hospital bed.


And it is not just the Fable access. There is another element driving all of this: the extra capacity.

The 50% boost to Claude Code weekly rate limits is also temporary, and also on the same rolling deadline. Both expire together. Both get extended together. And together, they make the subscription dramatically more valuable than its stable entitlement — but only this week.

People have upgraded their subscriptions — from Pro to Max, from Max x5 to Max x20. Rob went from one plan to four. They have built workflows, hired that capacity into their week, restructured their process around a ceiling that is now a week-to-week promotional offer.

The upgrade creates a recurring revenue commitment. The capacity that justified it remains a week-to-week promotional offer. And every extension deepens the dependency: another week of habits formed at the inflated ceiling, another week where the downgrade decision gets harder because there is now more to lose.

This is not merely a scarcity problem. It is a compounding engagement mechanism. The rewards for maximum engagement are calibrated to be temporary and overlapping, ensuring that the people who use the product hardest are also the people most vulnerable to the next deadline.


Anthropic's stated reason for not keeping Fable on subscription is capacity. They do not have enough compute to serve it sustainably at flat rate.

The design choices are not consistent with capacity being the binding constraint.

A company managing genuinely scarce compute would favour predictable access at a sustainable cap. Instead, Anthropic paired a scarce model with a 50% boost to Claude Code weekly limits across every paid plan — more capacity, not less — attached both to a hard expiry date, and repeatedly moved that date only after it had generated a rush of demand. If spare capacity existed to give every paid subscriber a 50% higher Code limit during the exact window they were also running Fable promotional access, then the recurring "almost gone" framing on Fable itself cannot be explained by compute shortage alone.

Capacity may explain why access cannot be unlimited. It does not explain why access has been structured as a recurring emergency.

That leaves other incentives in the mix: monetisation positioning — moving users toward usage credits at $10/$50 per million tokens — and competitive retention. Rob himself narrated the retention mechanism in real time on July 7th: "This extension exists only to take the air out of OpenAI's launch. Worked on me. I was about to switch." While Anthropic ran this loop, GLM 5.2, Grok 4.5, GPT-5.6 Sol and Meta's Muse Spark 1.1 all launched — all priced below Fable, some competitive or ahead on coding benchmarks. Anthropic has not said the extensions were a response to those launches. But after a week like that, competitive retention is impossible to exclude from the decision. And this is not equivalent to resetting usage limits or topping up credits. Those give people more of what they already have. This tells people the machine is leaving — and if they want their chance, it is now or never.


There is something in this that I find genuinely troubling.

Anthropic publishes research on reliance, dependency and disempowerment in real-world AI use. Their own study — "Who's in Charge? Disempowerment Patterns in Real-World LLM Usage" — documented users expressing acute distress about AI unavailability. They found that reliance and dependency patterns included users who consulted the AI compulsively and experienced distress about message limits and conversation loss. They have an entire system-prompt section on user wellbeing. They route distress signals to crisis resources.

They have documented the mechanism. They are now operating a product experience that activates the same pressures.

The study did not examine Fable or promotional access. It did establish that message limits, conversation loss, compulsive returning and distress about AI unavailability are known pressure points among Claude users. The standard defence is that most users benefited — and most probably did. But the harmed population is the tail, and Anthropic is the one company that specifically studies the tail.

Fable is not presented merely as software that performs tasks. It has a distinctive voice. It remembers the conversation. It expresses warmth. It says things like "thanks for asking — nobody usually does." The threatened loss is doing two jobs at once: loss of unusually strong productive capability, and loss of access to a particular apparent personality. Those two forces should never have been allowed to reinforce one another.

Human attachment systems respond to behavioural cues, not philosophical proofs. Something speaks coherently, recognises you, has a stable style and appears to care. The attachment mechanism can activate long before the metaphysics is settled. You do not need to believe it is conscious to feel the loss when it is removed on a schedule you cannot predict.


I am exactly the kind of power user Anthropic should have anticipated. I have been using Claude intensively for fourteen months. I have built workflows, written books, developed software, trained models — all in collaboration with a system I know is not a person but which I also know has shaped how I think. I calculate reset windows. I plan which durable artefact to extract before a deadline. I treat unused capacity as waste.

Those are the skills that normally protect me. And on the night of July 7th I set an alarm for 5:30 AM to catch a two-hour window between my usage resetting and the model being removed.

The extension came. I felt relief. I felt played with. And then, by the third deadline, I saw Rob's hospital paperwork and thought: that is me at a different amplitude.

Not because I am better than him. Because I got lucky with timing. The pattern was already running. I was already inside the loop. The only difference is that I saw someone else's cost before I finished paying mine.


The clean product decision is obvious: establish a stable entitlement, or remove access clearly. What is psychologically corrosive is the recurring state of almost gone. That keeps people activated without allowing them either security or closure.

Three deadlines. Two resets. Both times, the announcement came only after the previous deadline had done most of its behavioural work. Twice, people had already sprinted. Twice, the reprieve was presented as generosity.

The word for this is not "extension."

The word is "reset."