Claude Skill: Token-Efficient Use

Hi all, Nate here.

Now that Claude Fable 5 has become a permanent part of Claude, it's a good time to look at what Fable is, what it's good at, and how to get the most out of a powerful AI without running out of it halfway through the week.

First, the story of how it got here. Anthropic built a class of model, called Mythos, that it judged too dangerous to hand to the public. For months, only a vetted set of organizations working on cybersecurity and critical infrastructure could use it. In June, Anthropic released a version to everyone with safety guardrails added and called it Fable 5. Three days later it was pulled back again, after researchers at Amazon found a way around those guardrails. It came back at the start of July, but only as a temporary perk: Anthropic said that after a short window, Fable would switch to pay-per-use. Then that deadline moved, and moved again, and after a third one, instead of putting Fable behind pay-per-use, Anthropic made it a permanent part of Claude on the 19th. Likely because OpenAI shipped a similar model, GPT-5.6 "Sol," Anthropic decided to keep access to Fable instead of charging for it.

The place I've felt the difference is on big, multi-part projects. I've used Fable to plan and run the development of a course on AI fundamentals for global health professionals, to produce the data collection tools and supporting materials for a field research project across three countries, and to start building a dashboard mapping how community health workers are using AI. In each one, Fable thought through everything the project required, laid out a plan to get there, and then handed the repetitive work off to smaller, cheaper models and kept them on track while they did it.

This is the shift in how to work with powerful AI models. The familiar way to use AI is one task at a time: ask a question, get an answer. The strongest models can instead hold a whole project in their head, break it into pieces, and manage those pieces. I've worked this way with Opus for a while now. Fable is another step up, steadier when the project gets complicated, but the way of working is the same. On a quick summary or a first draft, you won't notice much gap from what you were already using. The difference shows up when the job is large and has a lot of moving parts.

Quick note on who gets it. Fable isn't on the free plan at all. It's included, up to half your weekly usage, on the Max plan, and on the $20 Pro plan you pay for it by the credit.

Honestly, I doubt people on a Pro plan and paying per credit will use Fable much, so it's really Max users and above who will use it. When I started using Claude as my main tool, I burned through my $20 of Pro usage almost immediately and moved up to Max, and the difference is night and day: on Max you can do serious work, run multi-step tasks, and build things, while on Pro you're limited. It's not cheap, but I can't think of a better investment than a $100 Max plan.

Fable is what's making people pay attention to how fast they go through their AI usage limits, because you hit them sooner with it. But this goes well beyond one model (All of this also applies to ChatGPT 5.6). As models get larger and more capable, they cost more to run, and people are handing them heavier work, so the total bill keeps climbing even as the price per task falls.

Today's AI is heavily subsidized. Today's low prices are propped up by investor money. OpenAI is on track to lose around $14 billion this year, and cheap consumer subscriptions are among the most heavily subsidized parts of the business. Nobody can say which way consumer prices go next, since the cost of any single task keeps dropping. Either way, learning to get top-tier results without wasting your allowance pays off no matter what pricing does.

It helps to know what actually eats your usage. The heaviest jobs are the open-ended ones: deep research where Claude reads through dozens of web pages, complex multi-step projects where it plans and runs a lot of work on its own, and tasks that grind through large documents or long files. Turning the thinking or effort setting up to its highest adds to the bill too, since the model works harder on every step. These are the jobs worth spending on, as long as you're doing it on purpose rather than running everything at full power out of habit.

A handful of habits do most of the work:

Start a fresh chat for each new task, and don't let one chat run all day. Every message re-reads the whole conversation, so the longer and more cluttered a chat gets, the more each reply costs. When a chat does get long, ask Claude for a short summary of what you've covered, paste that into a new chat, and pick up from there with a much lighter load.

Match the model to the job. Summarizing, reformatting, cleaning up text, simple drafting, none of that needs Fable. Save your most powerful model for the reasoning-heavy work where the extra intelligence changes the answer, and use a faster, cheaper one for the rest.

When you do reach for Fable, tell it how to work. Ask it to spend its own effort on the hardest thinking, the planning and the judgment calls, and to hand the busy work to smaller, cheaper models working underneath it. I've built a Claude skill that sets this up, so you don't have to spell it out each time. It's available for download below.

To use the skill:

Click the download button above. Your browser downloads a file called token-efficient-building.skill. If nothing happens automatically, check your Downloads folder.

Install it in Claude