(1st impressions) I burned an entire Claude Max 20x on Fable 5.1 in 8 hours

me when I type a post

tldr it's really fucking good, best model I've used to date. Expensive as hell tho. Writes well, I had only 1 experience with writing that was too dense to understand and had to ask it to deconstruct, other than that it still feels like good old Fable 5 just smarter. W model release.

So as you know we got a reset this morning

and my account reset was at 8pm today.

This gave me 3 windows (tail end of my 9-2pm + 5h + 2h) to burn all of it on Fable 5.1.

So I did. I also intentionally made my requests a bit demanding... usually even with models like Sol I usually spend a lot of time yapping until I'm satisfied I have completely defined the request. Here I was being super demanding like "Read issue 1621 and make a solution that aligns with the intent of this subsystem + make the PR" just to see how it would do. Knocked a whole bunch of backburner tasks and a few P1s out.

I dream of having unlimited tokens... was such a nice experience...

Day 1 take: Fable 5.1 is really good at interpreting intent and turning it into implementation, even better than Fable 5.

BIG CAVEAT: My primary repo is a well architected and deliberately designed system by a couple ex-faang friends and me, we're all from the pre-AI days, so the repo is already designed for meaningful work to be done with plenty of examples of good design and style that any agent can pick up on and hit the ground running.

Also some months back I pushed hard to early adopt the knowledge docs standards which have also helped greatly - A very up to date AGENTS.md, maintained around 39.8k characters (just below the 40k recommended limit), which references per-subsystem knowledge docs that I keep up to date on every PR. Less of a rules doc more like a living semantic map, which is what I find works best for agentic work for our use cases

So these results are tuned for a structured codebase rather than vibe coders though I can obviously extrapolate that it will do a nice job on vibed stuff as well

A few anonymized examples:

- I proposed adding last-accessed data to detect orphaned extract streams to save money in a data science subsystem. It traced every consumer and their specific quirks, proved that we can use our existing durable reference record system without any collisions, and implemented the design end-to-end without adding a redundant access-tracking system.

- I asked for a “dependency-graphed test determination thingy” to save money on github testrunner spend. It found the testrunner lib's existing related-test feature and built the missing CI change classifier around it, tested it, made the PR.

- I handed it a parallel-test flake and guessed teardown. It traced the actual failure to a sibling worker rolling back a shared test connection and then made a hook for that test file so future tests won't accidentally reintroduce the issue.

- I proposed putting a new capability on an abstract base class. It noticed that this contradicted the architecture’s language-neutral wire boundary and put the capability in the shared data envelope instead. Actually correct - it's what I would have done

Some numbers from the 8 hour window (sol copy paste sorry not sorry):

- 51 transcripts: 14 parent sessions + 37 subagents

- 1192 total turns (including a few Opus 5 subagents)

- 881 Fable 5.1 turns

- ~1.32M Fable output tokens, ~698k reasoning

- 164,190,733 total cached tokens

- 118,626,370 Fable 5.1 cached tokens

- Fable effort usage: 452 Max, 288 High, 101 XHigh, and 40 Medium turns