Since July 1 I’ve maxed out my Fable 5 usage three times and burned $172 of my $200 in usage credits, plus $174 upgrading to the 20x plan, to do it. Here’s what I learned about exactly how I’d use Fable 5 in the future.
If you’ve been under a rock, you may have missed the hullabaloo around Anthropic’s release of Fable 5. TLDR: they released it on June 9 and on June 12 the government restricted access to Fable 5 and Mythos. On July 1 Anthropic had to go ahead and re-release Fable 5 and in doing so they gave all of us subscription customers half of our plan usage in Fable.
Full disclosure, at the beginning of this adventure I was on the Claude Pro Max 5x plan, which has always been more than ample to do my week’s work, still staying under 95 of my usage so as not to burn extra usage credits. Fable 5 changed that. By Friday the third when my usage reset, I had already hit 93 on Fable and 83 on my plan. Between Friday and Sunday I did it again, taking my week usage to 85 and my Fable usage to 92. Late Sunday afternoon I upgraded to the 20x plan to finish off some of the work before the Fable 5 deadline, with a surprise that my Fable usage got reset to zero. Anthropic then extended access on all paid plans through July 12, and has since pushed it again to July 19, so the window is still open as I write this. By Thursday morning I was at 51 on Fable and 42 on my plan, with $172.04 of my $200 in usage credits burned. Then I stopped. After ten days of grinding Fable I needed a break, so the weekend went to friends instead. A week later the meter backs that up: 16 on Fable, 28 across all models, and the credit counter sitting still at $172.04.
What is Claude Fable 5 like to build with?
Here’s what I learned: Fable is an extremely capable model. It’s definitely a class above Opus 4.8 or anything else that I’ve used. I started off with Fable by having it orchestrate a review of a small task management web app that I had built earlier in the year. It’s a React-based app that I’ve been using mostly via a web interface for some task planning. It was my first experiment using Claude Code and I iterated it two or three times before eventually putting it to the side as other interests and builds came online. This seemed like a great place to see Fable flex its muscles. During the Fable blackout I had rebuilt the app in an overnight Sonnet and Opus run:
Converting it from a web app running natively on my system to a Docker container
Creating an MCP server for the app, so any AI assistant plugs into its functions directly (13 tools: tasks, timers, workload, channels)
Exposing it via Docker’s MCP gateway
Adding features that I liked from my preferred task management app:
A task timer
A workload counter comparing planned versus actual hours, with an over-commit warning
Time logging, so the app owns my actuals
Channels for work and personal contexts
A backlog and an interactive Today view
Google Calendar integration and AI planning
A seed importer for the 128 tasks from my old task manager
The moment access came back on July 1, I pointed Fable at it for a full code review and hardening pass.
This is the project where I spent most of my usage credits. I watched Fable start its run after the plan and blow right through my usage and deep into my usage credits. So learning number one: Fable is expensive. We all knew that but I felt that in my pocketbook.
The big question is: what did I learn from that? I learned that Fable is incredible at code. It caught so many defects not only in the original code but also in its own writing. Here’s how Fable structured the review: a diff pass first, then 10 finder agents fanned out in parallel, each hunting a different defect angle, 44 candidates found. Then one adversarial verifier per candidate, 28 running in parallel, each confirming or killing its finding, several by live reproduction. Then a fresh gap sweep. The output: 15 ranked, verified findings, the fixes, and a background agent standing up the Docker container to validate the MCP server end to end.
How should you structure a Fable 5 build?
From there I learned a big lesson. Plan with Fable. Let Fable do the initial review of the code. Chew on it. Give recommendations when it comes to execution. I learned for my next build to have Fable become the master agent. Use Opus agents as phase or stage orchestrators and use Sonnet 5 and Haiku agents to actually do the heavy lifting.
Here are some of the other things I had Fable 5 do while I was working on my second build:
Refactored 39 of my AIOS skills to a 10 out of 10 execution standard across 6 waves, with 74 cold-run fixes
Hardened my “Me Builder” prompt through 8 versions, including finding and closing a prompt injection gate
Interviewed me to lock my brand thesis, then assembled the full brand document
Interviewed me business-coach style and delivered the first draft of the business plan for the business my wife and I are starting
My third build was what I’d been waiting to do for months: completely refactor, rebuild, and augment my content production skill pipeline. Now I had dialed in the Fable architecture, using Fable for what it does best: deep thinking and analysis, and using Opus and Haiku as the execution engine. The numbers: 18 agents, roughly 1.6 million tokens, four phases shipped in a day. Six core skills refactored into brand-agnostic execution engines, with the brand supplied at runtime from a brand thesis document. I answered 5 questions the entire run and zero edits landed without my approval. Over the next two days the same architecture shipped a metrics skill wired live to Google Analytics, an SEO and discovery skill, and a 2.0 rebuild of the pipeline orchestrator. The verification wave caught the defect worth writing home about: a brand-new skill, built clean, never wired into the orchestrator. The lesson: a new skill isn’t shipped until its orchestrator calls it.
Is Fable 5 worth it?
So the question is, at the end of all this, is Fable 5 actually worth it? I would say it’s a resounding yes. Is it worth it to do everything? Absolutely not. I’ve learned that if I do use Fable 5 in the future, or any future Fable models at the price point that Anthropic is posting it at, I would use it very, very carefully like this.
How I prompt each model
From the Fable 5 guide from Anthropic, these are the things that I would add either to your Claude.md or in the prompt, or, in my case, I have a list that has all of these instructions for Claude based upon the model that’s reading it. Quick shout out to Karo Zieminski who wrote the best prompting article I’ve yet to read. Here’s a link to her article. The condensed framing below follows Charlie Hills’ summary of the official guide.
My per-model role assignments:
Fable: master agent. Deep thinking, planning, review, final judgment
Opus: phase and stage orchestrators
Sonnet 5: the working layer. Skills, spec checks, research, drafts
Haiku: the mechanical layer. File moves, greps, frontmatter sweeps
And the 12 behaviors from the guide: tell it when to act, have it check its own progress, set boundaries, only pause when needed, keep it running on long jobs, give it a memory, give it the reason behind the task, spin up parallel subagents, lead with the outcome, write for the absent reader, don’t over-engineer, and verify with a fresh subagent.
This is the first in a short series on what building with AI has taught me this year. Two more are coming: the six-month path from chat-user to architect of agentic systems, and why most of the AI advice in your feed is wrong.
If you’re running Fable 5, or weighing whether to, I want your read. What have you pointed it at that earned the price? And what did you pull back to a cheaper model? Tell me in the comments. I’m comparing notes.
Disclosure: This article was written by me, with light editing and polishing assistance from Claude.
You and A are starting a business? Super cool!! I have to be honest with you: I’ve been taking the summer off. The only thing I’ve been working on is my tan. But I have to say, it is spectacular. Love you two - B