Fable 5.2 is Anthropic's strongest model, and per token twice as expensive as Opus 5.5. Use it only where your current model falls short. For everything else it is too expensive.
What we came across
We follow a fixed circle of creators and newsletters who work with AI every day. This week six pieces about Fable 5.2 appeared. Per topic: what they say and what we think. We already wrote about the price per task on 11 November as well.
The price per token says little
Fable 5.2 costs 8 dollars per million input tokens and 40 dollars output, The Rundown AI reports. That is twice Opus 5.5, although it is 20 percent cheaper and up to 30 percent faster than version 5.1.
Our view: count per task, because a more thorough model uses more tokens and a task can end up costing more.
It shines where other work gets stuck
The creators who tested it mainly see gains in hard and visual work:
- Stronger at visual and design work, but slower and over 70 percent more expensive (Ben van Sprundel, on three of his own jobs).
- Unblocked a project that had been stuck for over a month, after three failed attempts with other models (Nate B. Jones).
- Nice demos of presentations and interactive explainers, partly made by others (The AI Advantage).
Our view: the stuck project convinces more than the demos, because it solved a problem that had been lying around for a month.
What it costs depends on how you pay
The two testers ended up at over 70 percent more and at nothing extra. A subscription comes out of your allowance, through the API you pay for every token.
- In his test Ben paid over 70 percent more than with Opus 5.5.
- Nate paid nothing extra, because it fell within his subscription.
- Nate does hear of 150 dollars for a single strategy document.
Our view: follow Nate's advice and first have it do one step with a cost estimate, because you do not know in advance which side you will end up on. Set a spending limit too.
Use it where a mistake costs money
Ben mentions work in which every figure has to be right, such as a financial model or report. Nate saw it succeed on a project that was stuck, with lots of context and earlier mistakes.
Our view: only use it where your current model demonstrably fails, because otherwise you mostly pay for nicer wording. So not for work that already goes well.
Choose up front and give the whole assignment
Ben goes through eight rules from the prompting guide Anthropic published for Fable 5.2. Three stand out:
- Choose your model before you start. Switching costs the reasoning built up so far and the cache.
- Give the whole assignment in one go: goal, rules and what "done" means.
- Clean up old instructions in your standing Claude files.
On the effort level he says different things in two videos: first high, later medium, based on Anthropic's guidelines.
Our view: start at medium, because a higher level costs more tokens and time. Only raise it when you can say why.
What to do now
Three steps to start with tomorrow:
- Make a lighter model the default. If you can describe the task precisely in one sentence, you do not need the heaviest model.
- Agree who uses the heavy model, and for what. With a spending limit attached.
- Test it on your own work. Compare result and cost on a few tasks you know well.
How it works under the hood
Model and effort level are two dials. The model determines what Claude knows. The effort level determines how many steps and checks it makes, and so how many tokens it costs. If the result disappoints, this rule of thumb from Ben helps:
- Did the model not know the subject? Choose a bigger model.
- Did it skip steps? Raise the effort level.
Switching models is expensive: each model has its own cache and reads the whole conversation again, at the full rate. The test figures in this piece come from the creators themselves; treat them as an indication.
This news touches chapter 9, 10 and 24 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to publications by BenAI on prompting, The Rundown AI, Nate B Jones on YouTube, The AI Advantage, Nate's Newsletter and BenAI on model and effort.
S.O.S