Stop Fable 5.1 from wasting tokens

Fable 5.1 is a beast. And it burns tokens like one.

Here is how to cut token waste with four commands, and free open-source repos that cut it even further.

None of this touches output quality. It changes what you pay for the same result.

And it all works whether you're on Fable 5.1 or not, because token cost adds up the same way regardless of which model you run.

Now let's start.

Step 1: Drop your effort level.

This is the single biggest trick, and it's a setting you can control on every request.

There are five levels: low, medium, high, xhigh, and max.

Effort controls how much the model reasons before it answers. Higher effort means more thinking before a response, which costs more tokens whether or not the task actually needed the thinking.

Here's the way to think about it. Effort level is how long you let someone think before they answer. Ask anyone what 2+2 is and they say "4." Ask them to think hard for ten minutes first, and you pay for ten minutes of thinking to get the same "4."

Well...most tasks are 2+2.

Anthropic's guidance for Fable 5.1 is: start at high, which is the default. Step up to xhigh or max only for the most capability-sensitive coding and agentic work. Drop to medium or low for routine or latency-sensitive tasks, once the lower level is confirmed to still hold quality.

The numbers are wild. On CursorBench, Fable 5.1 at low effort scored higher than Fable 5 at high effort, at a third of the cost.

Test it before trusting it. Pick a task where the right answer is already known, then run it low.

"Run this request at low effort and tell me what changed in the response"

Use one of your real tasks. Compare the output directly before assuming low effort means worse results. Keep the heavy stuff, multi-step reasoning, real debugging, anything where a wrong answer costs time later, at high or above.

Step 2: Run cost-optimize.

This one launched August 26 as part of the claude-api skill. It looks into recent usage and ranks the cost levers that matter for a specific project.

1/claude-api cost-optimize

It checks caching, token hygiene, batch processing, effort level, and model choice, in that order. Caching and token hygiene are usually the cheapest fixes and pay off the fastest, before it ever suggests something structural like switching models.

Each suggestion gets approved or skipped one at a time. Nothing changes without confirmation, so there's no risk of it rewriting a setup.

Step 3: Run prompt-audit.

Prompts and skills collect junk over time.

Think of it like a junk drawer. Every edit puts one more thing in, and you don't ever clean it out. Except this "drawer" charges you tokens for every request, forever, until the junk gets removed.

1/claude-api prompt-audit

Run it on any Skills folder you have. The old skills, the ones that you've probably edited over and over, are exactly where it finds the most waste. Every edit added something without removing what it replaced. Run this and you'll notice the difference.

Step 4: Migrate old API configs.

If a project is still running configuration built for an earlier model, this handles the model ID swap plus any breaking parameter changes across the codebase.

1/claude-api migrate this project to claude-fable-5-1

Name the actual target model. It confirms scope first, the whole working directory, one subdirectory, or a specific list of files, before it edits anything. Then it gives you back a checklist of whatever's left to verify by hand.

Four repos to cut waste even further.

These aren't specific to Fable 5.1. On any model, the same token saving applies, so it's good to know what's out there and what each one actually does.

Caveman

Compresses the model's own replies into short, telegraphic responses instead of verbose ones.

Setup:

1curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

RTK

A Rust proxy that compresses shell command output before it reaches the model's context.

Setup:

1brew install rtk
2rtk init -g

Ponytail

Makes the agent write less code in the first place. A senior-developer lens that picks one line over fifty.

Setup:

1/plugin marketplace add DietrichGebert/ponytail
2/plugin install ponytail@ponytail

CodeGraph

This one works completely differently. Instead of compressing text, it builds a knowledge graph of the codebase, so the agent can query symbol relationships and call graphs directly rather than scanning files with grep and read.

Setup:

1npx @colbymchenry/codegraph
2cd your-project
3codegraph init -i

Where to start.

If nothing else gets done, do Step 1. Drop the effort level on the next routine task and compare the output. That one setting does more for your credits than any repo, and it costs nothing to try.

Then run cost-optimize and prompt-audit on any of your projects.

The four confirmed settings from Anthropic's own team beat anything installed from a repo. Go looking at the wider ecosystem after that with the four repos I gave you.

If this saved you tokens/money, make sure to like and share it with your network.

See you in the next one :)