The Smartest Model Is Best Not Used for the Work
After Claude Fable 5.2 launched, one early user offered a counterintuitive way to use it: let the strongest model do only planning, and leave execution to cheaper models. This division of labor is becoming a default practice.
Anthropic released Claude Fable 5.2 and Mythos 5.2. The official description is “built to tackle the toughest technical challenges,” with the price dropping from $10 to $8 per million input tokens and from $50 to $40 for output.
But more interesting than the release itself is the usage suggested by early user Alex Finn. After using them for a few days, he recommends the opposite: don’t let Fable 5.2 do the work.
His workflow is like this:
Use Fable 5.2 for all planning and brainstorming
Go back and forth until the idea is solid enough
Have it break the task down in extreme detail and write it into Linear or Notion
Switch to Opus 5.5
Have Opus 5.5 execute the tasks one by one
Use Sonnet or even Haiku subagents for testing
His exact words were: “Fable 5.2 is the smartest model ever. But if you have it break down the task in enough detail and plan thoroughly enough, you don’t need it to do the execution.”
The logic isn’t complicated. Fable 5.2 is priced several times higher than Opus 5.5; use the most expensive brain only to think, and leave the hands to cheaper ones. Alex says this setup “produces better results and saves a large number of tokens.”
Official data also supports this layered approach. On FrontierCode 2.0, Fable 5.2 with High effort scores about the same as Opus 5.5 with Max effort, but costs about half as much per task. At Xhigh effort, it gets 49.1%, the highest among all released models.
Official recommendations follow the same logic: for everyday coding and knowledge work, start with Opus 5.5; when encountering complex tasks, move up to Fable or Mythos.
Some demonstrated results also support this positioning. 1,000 Fable 5.2 agents completed the formalization of Fermat’s Last Theorem within 30 hours, while Opus 5.5 with the same setup advanced only 26% in 18 hours.
For AAV capsid structure prediction, Mythos 5.2 accurately modeled 28 out of 34. On the RoboDojo simulated dual-arm robot benchmark, Fable 5.2 scored 64.9%, more than twice GPT-6 Astra’s score, and far above Fable 5.1’s 27.6%.
Still, the comments were not without objections. A user named Peter said: “Fable 5.2 isn’t worth using; its performance is basically Opus 5.5, and it will burn through your usage limit.” Another user asked directly: “Will this burn through my weekly token limit for my large repo?”
So the truly awkward thing about Fable 5.2 is this: it is the strongest, but not the most cost-effective. And Alex’s division of labor is a solution offered after acknowledging this fact.
Someone summed it up even more directly: “Planning model plus cheap executor—every team does this. What’s new is that the executor got weakened mid-sprint.”
There is also feedback from a non-programmer user worth noting: “For non-programmers, planning being the bulk of it is everything. I can’t rescue bad execution, so how detailed the task breakdown is is decisive. In a ninety-minute evening window, most of it goes to planning, and there’s almost no time to watch it run.”
This is probably the most practical change in this round of Fable 5.2’s release. It isn’t that there is yet another stronger model you can use mindlessly; rather, you have to start thinking carefully about which step should use which brain.
发布时间: 2026-11-10 07:32