Fable 5 Loops: Why Anthropic Wants You Burning Tokens

The skill that matters with Fable 5 is token allocation, and token allocation is just capital allocation: what deserves tokens, what deserves the best model, and what deserves your attention at all. Everyone calls themselves a CEO now; this is the part of the job they skip.

The loop pitch reads differently once you line up three events. Anthropic is heading toward an IPO. Boris Cherny, the creator of Claude Code, is on podcasts saying the job now is to write as many loops as possible so thousands of agents run in the background. And Fable 5 launched inside the subscription, with a switch to metered compute credits two weeks later, on June 22. The IPO needs growth, loops maximize burn, metered credits monetize the burn, and the one paying for it is you.

The slides for this talk are on the site as well.

Is the loop pitch real or a sales ploy?

I am not saying loops do not work. I am saying it is a very easy thing to evangelize when you can use as many tokens as you want and never think about the economics, and that it happens to be exactly what a company approaching public-market scrutiny needs its users doing.

We have seen the playbook. Ubers used to cost $5; the IPO ended that. LLM subscriptions are heavily subsidized today, and providers are moving toward token-based billing. Fable 5 itself points the same way: it is more token intensive than Opus and it is very good at long-running tasks, which is another sign of work moving into the background, where the meter spins without you watching it.

What loops actually do in production

Thousands of agents fixing bugs and submitting PRs overnight sounds great. From my experience in production it mostly either does not work or produces busy work. There are maybe three valid use cases where I want something running autonomously in the background.

The reason is drift. A loop is unmonitored, so your review becomes the endpoint, after all the drift has already happened. I regularly find myself redoing background work because it did something dumb, and the error compounds the longer the run. That is twice the token spend for one result. Three things in parallel is slower than one thing you ship front to end, especially once you count the review.

If your requirements are loose enough that you can live with whatever comes out, loops are fine. If things need to work properly, run the hard tasks yourself and look at the output.

Tokens are getting more expensive, not cheaper

The promise was that this becomes electricity: so cheap nobody cares. That might happen eventually. What is actually happening is the opposite. Fable 5 uses twice the usage quota of Opus and costs twice as much via the API, and since everyone insists on the cutting-edge model, you are always running the most expensive one.

Enterprise accounts are already on token-based billing, and plenty have cut seats because nobody is trained to be discerning about usage yet. The $20 subscription creates an illusion about what compute costs: ask someone to pay for a job that reads 50,000 files and they ask why it is not included. My own usage on the $200 Max plan works out to something like $5,000 in API credits. The gap is massive, and it will not stay subsidized forever.

The multitasking tax

The loop pitch also quietly bills you a second time, in attention. If your agents create a pile of PRs overnight, who reviews them? Merging blind is feasible when your only user is your mom, not when you run a production system. I have seen Claude Code do stupid things for no reason; I do not trust any of it without review. Context switching across parallel background work is a tax on the scarcest resource you have.

What I do instead: one ticket at a time

My system changes every few weeks because the tools change that fast, but the shape holds:

1. Write the ticket by hand. Three bullet points beat fighting an AI-written ticket that sneaks in fifteen features nobody asked for.
2. Triage the effort myself, sometimes with Claude assessing the blast radius: how many files change. Mostly it is gut feeling from twelve years of writing software.
3. Anything under a three in effort goes to a background agent, like Claude's GitHub app.
4. Everything above that I do myself, one ticket at a time.

Most loop output moves the needle only for the provider, which is, surprise, the party that wants to IPO. Since June 22 the meter is running, and it will show you exactly which of your loops were worth anything.

If you do run background work, the mechanics matter: fork agents are how delegation works without losing your session's context, and the same small-task rule applies there too.