Burning the 3:40 AM Window | Kevin London

(history.length === 1) ? window.location = '/' : history.back())()" data-cf-modified-87ccb287c29f3fe7d8eb5cc7-=""> Go back<br>Burning the 3:40 AM Window<br>Published:Jul 7, 2026<br>| 5 min read<br>On Anthropic’s Fable 5 release, a friend ran an audit of his codebase. Fable spun up 101 agents and, after 30 minutes, exhausted his five-hour subscription window. The sub-agents stopped when usage ran out, and he lost the work. He couldn’t tell if Fable was better than Opus because, from his perspective, all it did was burn his tokens.

Shortly after Fable’s launch, I heard the same discussion in three different places. Someone blew through their five-hour session limit with four hours still to go. Someone else asked if the session limit had changed.

Nope! It’s just a new model and it eats more tokens.

Anthropic says Fable is 2x the cost of Opus 4.8, but in practice it feels more like 4-6x usage.

I knew I’d need to write a post saying “hey if you use Fable on max effort for everything you’re going to incinerate all your tokens.”

Here is that post.

Unfortunately, Fable’s very good

The biggest difference I see between Opus and Fable is that Fable is far more capable of taking a big picture view, like collaborating with a staff engineer instead of a mid-level engineer.

I had an issue I couldn’t debug. I kept hitting dead ends, and tried four different hypotheses unsuccessfully with Opus. After an hour or two, I handed it off to Fable with a new hypothesis. Fable worked for 10 minutes and came back with “The 5th hypothesis is wrong too, and here’s what it actually is” along with a pull request to fix it.

Another time, I burned through my usage window with Fable and swapped to Codex. I designed and implemented a new feature with Codex, and came back to Fable once my usage window reset. Fable looked at the Codex work without knowing it was Codex, and said “this is all wrong, let’s build it this way instead.”

My friend with the 101 agents? He said when Fable finally completed the audit, it identified 30 problems and fixed them, including some funky edge cases around a billing module.

But it is expensive

According to ccusage stats, I’ve had several days where I would’ve spent over $1K in tokens at API pricing.

Is Fable worth $1K a day? Probably not.

But it is worth it for some things. I’d pay $30 to solve a problem I’d been working on for 2-3 hours, or formalize an API contract between teams, or optimize performance in a critical path.

Someone changed a slider from using notches of 0.02 to 0.1, co-authored by Fable 5. That’s probably not worth it.

The 3:40 AM window

When I first started using Claude in subscription mode, I’d try to maximize my usage. A friend tokenmaxxed by waking up in the middle of the night to use a new five-hour window that had just reset, burn the window, and go back to bed. The inventor nap schedule but for a billing cycle.

The morning I started on this post, I had a 3:40 AM to 8:40 AM usage window. I planned what to delegate to the agent, and woke up at 7 AM to feed Claude tasks, like it was a parking meter.

Rationing is back

A friend stretched their burn rate from 90 minutes to 3-5 hours by adjusting effort level per task. Fable 5 on low effort is surprisingly good!

I’ve stopped resuming 200K+ token sessions in Fable because the first message on cold, uncached context can burn 20% of a five-hour window. And the cache is only good for an hour. If you go away to lunch and come back to a Fable session, that’s new session time.

If I have tokens in my subscription, I want to spend them. If I was on API pricing, I might be more economical. In a way, it reminds me of when MoviePass came out with essentially unlimited movies. I knew someone who went to the movies every day for a month.

But that can’t last forever.

Today was going to be the day

In the beginning, my advice was “go full bore until Anthropic pulls it from the subscription.” Now, I’m only using Fable maybe 25% of the time, with 75% Opus because otherwise I’d be waiting for my window to reset. I think of using Fable like spending genie wishes.

Beyond the normal usage window, there’s a weekly Fable allocation. My weekly limit reset on July 6th at 1 AM and I thought I’d have to use my full Fable allocation by July 7th.

On the 6th, I woke up at 2 AM and considered giving my agent some things to do, and ultimately decided to go back to sleep like a normal person instead. I spent much of the day queueing up designs, reviewing my backlog, and looking for improvements, just to make sure I could use it on something before it went away. Then I saw it was extended. Now only 14% remains of my Fable usage for the week.

That makes me sad. I’ve grown fond of Fable.

I’ll have to pick my last wishes carefully.

Want to hear when I post?<br>Subscribe

ai

engineering<br>Back to Top Share this post on:

Submit this post to Hacker News

Share this post on X

Share this post on Bluesky

Submit this post to...