Article
Fable 5 vs Opus 4.8: the cheaper model won
The same task. Two models. Fable 5 against Opus 4.8.
Who won? Opus. It handled the task with a single round of checking for 721,000 tokens, while Fable needed nine rounds and burnt through 2.78 million tokens.
On paper Fable 5 is the better model, so why did it lose?
"It's not the plane, it's the pilot."
The better prompt won. And I know it, because I measured it.
Fable 5 went in circles and burnt almost 4x more tokens.
Nine rounds
I commissioned a detailed audit of how data is processed on one project. It was a document full of hard facts: how many times an hour something runs, when it runs, which line of code it sits on. More than a hundred such claims.
And here is the catch. Not one of those numbers can be guessed. Either the right file is opened and what is actually in it gets read, or it does not. An invented fact looks just as convincing on paper as a true one. The difference shows only when someone actually checks.
One version of the report was prepared by Fable 5. To catch the mistakes, I put checkers on it. Not people, more AI. They had a single task: find an error in the report. When they found one, it was fixed and the checkers went to work again. And again. The rule was: repeat until one whole round finds not a single error.
It took nine rounds before they delivered a result without errors. Thirty-six agents in total.
Why nine?
If I have the best model on the market, why did it take nine rounds?
The answer has three parts.
First: most of those rounds did not fix errors from the audit. They fixed the checkers' own errors. The AI that was meant to find mistakes made more of them than it found. Twice it reported that a certain process runs every five minutes. In reality it ran far less often. Twice it miscalculated the size of the data. Once it reported an error in a file it was looking for in the wrong place.
When the rule is "repeat until it is flawless" and those doing the searching are often wrong themselves, the last rounds are not spent fixing. They are spent proving that the alarm was false.
Fable 5 was not disciplined, it cut corners
Second: this task was not about intelligence. It was about discipline in checking. The answers were not in the model's head, they were in the files. And even the cleverest model will not tell you what is in a file until it opens it and reads it. The model was not disciplined.
Third. Down to prompting skill.
Check before you write
The core is one sentence. Fact-checking belongs at the input, not the output.
Fable 5 did it like this: a subagent finds a fact, Fable takes it and writes it into the document. Verification came only at the end, on the finished text. It should have been the other way round. A fact is verified in the file before it ever reaches the document.
A plain table was missing too: each claim, its source, and who verified it. Without it, the checkers in every round looked at different parts of the document at random, so nothing was ever covered all at once. The last unverified claims therefore took the longest. With a table, one honest pass would cover everything and it would be done. Without it, it was a lottery.
It is a missing rule about where and when checking happens. And anyone can set that rule when they write the prompt.
What those nine rounds actually cost
I have exact figures from each round.
Join the Library
Full access to my thoughts, personal stories, findings, and what I learn from the people I meet.
Join the Library · €29.99 per yearSummary
Common questions on this article's topic
Which AI model is better, Fable 5 or Opus 4.8?
Why does an AI make so many mistakes when checking facts?
How do you write a better prompt for an AI audit or review?
Does the model or the prompt matter more with AI?
How many tokens does an AI audit use?
What does checking facts at the input rather than the output mean?
If you have any thoughts, questions, or feedback, feel free to drop me a message at mail@richardgolian.com.
LinkedIn
Related articles
A few weeks ago I installed a small local AI model on my laptop that watches a live camera feed. I turned the webcam on in the dark, and in near total darkness it recognised me and the objects in the room. That such things exist, I have known for a long time. What opened my eyes was the accessibility. I installed it in one prompt, free, and it runs entirely on my machine, sending data nowhere.
I once wrote about building my own privacy-friendly analytics tool. It had bot detection from the first version, yet it was not enough. Direct visits took a strangely high share of my traffic. When someone claims that 20% of their visits are bots and 80% are humans, I used to think the same. Today I would say the opposite ratio is closer to the truth. This is how I got there.
I will speak for myself, for what makes sense to me to read and to write. In the age of artificial intelligence it makes no sense to me to be a secondary source, to summarise or comment on what someone has already written. What interests me is being a primary source: to do something, to find something out, to have a conversation, and to write about it. Because no machine and no algorithm can replace a unique experience.
More articles
I have Heidegger and my notebook beside me. I am asking where all of this is heading, where artificial intelligence is taking us.
Seventy per cent. That is where the first AI output begins, even when you give it the full company context and the best examples from the past. We are talking about the kind of output that cannot be defined programmatically. It is more complex. Often it is creative work. On one repeated type of output I reached eighty per cent within a week. Every further percentage point is harder than the one before.
For a long time we treated the internet as the main road. The place where work and relationships happen. Yet most of what we see on it today is, or soon will be, AI-generated: text, images, profiles and comments. The internet is turning into an online game full of bots, where you cannot be sure that a human is on the other side of anything. So I ask: was the online world the main road, or only a temporary detour that part of us will return from, back offline?
A few days ago I interviewed a senior marketer. An experienced man, years of practice. I asked him about AI. He said he barely uses it. He had one bad experience with the output and decided he was too senior for it to add value when it is not perfect. I know the other side too: professionals who automate everything that can be automated.
Europe does not have the capacity to face a full-scale, mass drone war of the kind we see in Ukraine. Three dependencies weaken it: China supplies the physical material for defence systems, the United States supplies capabilities Europe does not have, and twenty-seven states cannot agree how fast, or who pays. Rearmament plans exist, but they are being carried out slowly.
AI produces the graphic, the newsletter and the product page faster than a person. What is left for the one who used to do it is the judgement, knowing whether the output is good. But most people have worse judgement than AI. And whoever cannot judge quality cannot delegate either. How do you tell whether yours is the judgement a company relies on, or the kind it can replace?
In April, in the first part of this series, I wrote about an AI prediction system I had started building on my own machine. At the time the software was a few hours old and the prediction record was empty. The record since then has shown one thing: the system does not yet understand the market it is being asked to forecast. It can pull macro context, book value, earnings. But it cannot put those together into something that helps it understand the price.
Prague, 13 May 2026. On my way to work I started thinking about something that stayed with me for days. If most routine work on a computer disappears in the next ten years, and a large share of repetitive manual work disappears with it, what happens to the flow of money? Who pays whom for what? Which economic layers will exist, how large will they be, and what relationships will run between them? This is the six-layer map I sketched as an answer.
I am building an AI system to predict the S&P 500. It runs on my own machine, uses free public data (yfinance, FRED, the Shiller dataset), and grades every forecast against reality. This series documents the build itself: the decisions, the methodology, the mistakes. What I will eventually share from the running system is a separate question, and an honest one.
Four days in Catalonia. No computer, no AI, almost no social media. I bought this notebook so that I could write down what I would think about, and what I would come across and learn on the trip.