Original video in Chinese.
Key Takeaway
- Fable 5 is currently the strongest model ordinary people can use: it performs extremely well on “high cognitive load tasks,” including long context, multiple files, multiple tools, multi-stage reasoning, and complex Agent workflows. It clearly outperforms Opus 4.8 in task decomposition, Skill selection and invocation, and sub-Agent orchestration.
- Real-world validation: it can digest 200+ articles at once and distill a complete theoretical system (a concept table plus a programmatic document); it can review a 70,000–80,000-character manuscript and identify structural and long-context-level problems. It is well suited to serve as the “brain” of multi-Agent systems like newtype OS.
- Barriers to use and core values: the cost is relatively high (a $20 test budget was basically used up), and it runs slowly. It must be accessed through OpenRouter plus an overseas credit card. I emphasize solving infrastructure problems first, then understanding model capabilities correctly, and finally using it to build your own system rather than amplifying someone else’s system. That is what truly creates the gap between people.
Right now, the best model ordinary people can use is definitely Fable 5. After trying it recently, I was genuinely blown away. No wonder the U.S. government previously asked Anthropic to restrict access to this model on national security and export control grounds.
In one sentence, Fable 5’s strength is this: it is exceptionally well-suited to “high cognitive load tasks.”
What counts as a “high cognitive load task”?
Long context, long chains, multiple files, multiple tools, multi-stage reasoning, and complex engineering Agent workflows.
For example, one of our newtype Planet friends said that he used the Fable model to rebuild the company’s entire business platform, including the front end and back end, and got it done in an hour, deployed directly, with no bugs at all. Then he threw in two documents and finished the mini program and backend system in 40 minutes, also in one pass.
That’s what “high cognitive load tasks” are. In everyday language, that means the big jobs.
I also ran two tests.
The first was to have Fable read all of my past two or three hundred articles, then distill my theoretical system.
I had actually been working on this for more than a month, but I always felt the result was still a little off. Once I got Fable involved, I immediately noticed how it differed from Opus:
In task understanding and decomposition, in the selection and invocation of Skills, and in assigning tasks to sub-Agents, Fable is a full tier above Opus. Its logic and structure are extremely clear, and it is a pleasure to read.
After Fable’s diagnosis, my theoretical system did have a skeleton, but it also had problems like too many concepts and layers that were too scattered.
So I had it first generate a concept table, meaning what specific concepts I had discussed over time, such as worldview, values, methodology, and so on. Only with that table could I see which concepts were core, which ones conflicted, and so on. Then, based on the concept table, I had it generate a programmatic document.
I made 200 episodes of video. On the surface, they are about AI, but what matters more is what I want to achieve through AI, and why I want to do it. This programmatic document is a complete expression of those ideas. It is also the most important work I’ve been doing recently.
What Opus could not do, Fable handled in one go. It cost a little under $10. If you don’t understand how important that is, you might think it’s expensive. But in my view, I’d still be willing to pay $100 for it.
The second test was to have Fable review my manuscript.
I’m currently writing a new book called Using Codex to Do Self-Media. It’s a practical guide, around 70,000–80,000 characters long. The word count isn’t huge, but as you write, it’s easy to make mistakes.
Fable is very good at handling this kind of long-context work. It quickly found several issues, such as the opening promises not matching the actual content, problems with the structure of certain chapters, and so on.
To be honest, after every three or four chapters, I would have Codex help me review them. But the issues it found were all relatively superficial. Comparing that with Fable now, you can see that for problems that require traversing long context to identify, such as structural ones, GPT-5.5 and Opus 4.8 are both not very easy to get right.
So, as I said in the Planet: Fable is very suitable to serve as the brain driving newtype OS. Because this system has multi-Agent orchestration, a pile of Skills that need to be selected and invoked, and a huge number of reference documents—this is truly a textbook “high cognitive load task.”
Of course, using Fable does come at a relatively high cost. I topped up $20, and the two test tasks I just mentioned almost used it all up. Besides the money, Fable runs even slower than Opus.
Even if you can accept all that, there is still the problem of how to access this model. I used OpenRouter to call it, and you have to recharge with an overseas credit card.
That is why I keep reminding everyone in the Planet that you absolutely must thoroughly solve the infrastructure problem.
You know, in the AI era, how do the gaps between people widen?
Infrastructure is the first hurdle. It looks like a practical issue, but underneath it is really an issue of mindset—many people do not think this is important and believe that making do with a cheaper model is fine. That is a gap in mindset.
Once you get past this hurdle, the next thing is your understanding of models and various tools. If you know Fable’s strengths and you also know the cost of accessing it, then you know where to use it and how to maximize its value.
Then comes the final hurdle: what exactly are you going to do with this top-spec brain? Of course you can use it to finish company tasks and get things done quickly and beautifully, and there is nothing wrong with that. But you have to see clearly: if you do that, what you are amplifying is someone else’s system. Only when you use it to build your own system and accumulate your own assets are you amplifying and building yourself.
The first two hurdles are just the admission ticket; the last one is what determines victory or defeat. That is how the gap between people is created.
OK, that’s all for this episode. If you want to understand AI, want to reclaim individual sovereignty, and want to find like-minded people, come join our newtype community. See you next time!