Anthropic's Fable 5 model costs twice as much as its predecessor Opus 4.8, and while possessing significantly higher intelligence, it tends to "overdo it," burning extra tokens. However, as recent observations show, there is an effective task distribution system that can reduce costs by more than 50%. This method, known as 10-80-10, is actively used by Anthropic's own engineers.

The key idea of the approach is to use Fable 5 only where its unique level of intelligence is truly necessary. The rest of the work should be delegated to cheaper models. This is a kind of "cheat code" for using Fable 5 without the risk of going broke.

The Essence of the 10-80-10 System

The first 10% of the work is allocated to planning. Fable 5 acts as the architect: it defines the structure, approach, success criteria, and constraints before the actual work begins. If the plan is correct, everything else is built on a solid foundation.

The main 80% is execution. Routine tasks, iterations, and minor edits are handed over to cheaper models: Opus 4.8 for standard tasks, Haiku for light ones, and for some cases even GPT-5.5 is suitable. It is at this stage that the most tokens are burned, so paying double the price for Fable here is simply not cost-effective.

The final 10% is verification. Fable 5 returns to check the result against the original architecture: does it match the plan and are there any gaps? Since this is a check of a finished result, not generation from scratch, this pass consumes noticeably fewer tokens.

Instead of the usual "request — response — check — re-request" cycle, where the user themselves plays the role of the loop, it is suggested to set a goal once with the /goal command and configure an automatic cycle via /loop. After that, the model itself launches subtasks.

How to Further Reduce Costs

In addition to the 10-80-10 system, there are other practical tips. You should start not with the maximum effort level, but with the medium one: as data shows, Fable on medium mode outperforms Opus on high mode.

It is also recommended to delete old skills created for previous models, as shorter instructions are cheaper. Additionally, it is useful to give models not only the task but also its reason — this reduces the number of edits and, consequently, token consumption. The /usage command helps track spending.

Two costly mistakes are highlighted. The first is accidentally launching Fable: Claude Code and the app now use this model by default, so the model selector should be checked every session. The second is the absence of a spending limit. Fable will soon switch to pay-per-token, so it is worth setting a hard limit in the settings in advance.

Analyst's opinion: The 10-80-10 system is not just about savings, but a fundamental shift in the approach to using AI. It turns Fable 5 from an expensive "hammer for every nail" into a strategic asset that is used only for the most complex and critical tasks. In a world where computing costs are rising, such optimization becomes not just desirable, but critically necessary for any serious project.