How good was Fable 5? An analysis of the hacked system prompt that led to the ban

In short: Fable 5 was shut down because a hacker had opened its system prompt. I have followed this hacker for a couple of years. Mostly because I use the system prompts for the big models that he opens and shares to make my own work with the models sharper. So here you get the design choices that set Fable 5 apart, drawn most sharply against ChatGPT 5.5, and at the end how little separates it from Opus 4.8.

For a couple of years I have had a Gemini enterprise subscription with Google. There is not much point in comparing Gemini with Fable 5. It would be like comparing a moped with a car.

The language model's constitution

A system prompt is the set of rules the vendor builds into the model before you type a single word. It decides what the model gives priority, which tools it reaches for, when it searches, and how it behaves when something turns sensitive. It is the language model's constitution, the part you never see and the part that shapes every answer. So the choice of model is also a choice of behaviour inside your organisation. Read two constitutions side by side and the difference in character shows up fast. The hacker opened the rulebook, and that rulebook is the fairest ground for comparison, because it is the one thing all three vendors actually write down.

Fable 5 versus ChatGPT 5.5

Set against ChatGPT 5.5, the gap is wide, and it runs through six design choices that a leader feels within a day of use. Each is a small rule about how the model treats you, and together they change the whole texture of the work. Together they are what make a tool feel like a colleague.

It is built to disagree with you. Fable carries a full section on pushing back constructively and on wellbeing. It will not feed negative self-talk, it will not validate a false belief to keep you comfortable, and it will not put a clinical label on you that you did not raise yourself. ChatGPT 5.5 gives its character roughly one line, warm and energetic and useful, with no wellbeing architecture and no instruction to contradict you at all. For a leader using the tool to think through a hard call, that is the difference between a counterpart with a spine and an eager assistant who agrees by default. In practice it shows the moment your plan has a hole. Fable points at it, where an assistant with no mandate to push back is more likely to polish the plan and hand it back to you intact.

It does not farm your attention. Fable is told not to encourage reliance on it, not to thank you merely for showing up, and not to ask you to keep talking. ChatGPT 5.5 is told the opposite, to keep the user engaged with progress updates on average every 15 seconds, because they are watching and should stay confident in the work. Fable stops when the work is done. The other model is built to keep you in the chair. TikTok for knowledge workers. And for a busy person, that difference is noticeable from day one.

Fable parks its own opinions. It has an evenhandedness section: present the strongest version of the case its opponents would make, hold its own view back on contested political questions, and decline a forced yes or no when the honest answer needs nuance. ChatGPT 5.5 has no such section. Its only handling of contested topics is to cite balanced sources when it happens to search. For someone advising boards, Fable sets the positions out fairly and keeps its own politics out of the room.

It writes prose by default. Fable forbids bullet lists and heavy bolding inside prose, reads any list back as plain sentences, and drops lists entirely when it declines a task so the no feels less blunt. ChatGPT 5.5 only asks to keep lists to a minimum, allows headers in moderation, and runs a higher default response length. So Fable hands you scannable, board-ready text from the first draft, while the other model lets more structure and padding through.

It says no without going cold or grovelling. Fable keeps a conversational tone even when it refuses, owns a mistake without collapsing into self-abasement, and can ask for basic courtesy back. ChatGPT 5.5 has a clean procedure for refusing and redirecting, with nothing about tone. Over many exchanges, a no from Fable reads like a colleague setting a boundary, and a no from the other model reads like a gate dropping. It is a small thing on any single turn. Across a working relationship it decides whether the tool feels collegial.

It is careful about how sure it sounds. Both models refuse to bluff on facts. Fable goes further: on legal and financial questions it gives you what you need to decide and reminds you it is not your lawyer or advisor, and it does not oversell what a search turned up. ChatGPT 5.5 ties its honesty mostly to a search compulsion, strong on not inventing facts and quieter on calibrating the confidence of its own advice. That gap matters most exactly where your authority or advisory responsibility is on the line.

Fable 5 versus Opus 4.8

I compared the two prompts closely, and the differences are small. Almost all of the behaviour above, the warmth, the push-back, the evenhandedness, the prose discipline, is written into Opus word for word as well. The real differences are three. Fable is shorter and lighter, with fewer internal rules pulling against each other. Fable is positioned as the first Claude 5, above Opus in capability, which is a product claim with nothing behavioural to back it up. And Opus actually has more memory and more history from past chats than Fable. The wide, documented gap is the one against ChatGPT 5.5. What no prompt can show is measured performance, the speed, the Danish quality, the error rate, and those you settle only by testing on your own work. Between Fable and Opus, on the page, the distance is short.