Lennox Saint
Open menu
LAB 0040

on YouTube 12 nov 2026

field notes · new episode

Claude Fable 5.2 just dropped. here's the one job i'd actually pay it for

from the episode: Claude Fable 5.2 just dropped. here's the one job i'd actually pay it for | LAB 0040

Lenny's AI Builders - LAB 0040

the boss and workers setup, one prompt to test two models blind, and the Fable reset catch.

Claude Fable 5.2 just dropped and there's only one task i'd actually pay it for.

the one task is the boss. it reads the task, picks which worker gets it and knows when it's actually good. Opus 5.5 already does the work.

so i ran Fable 5.2 against Opus 5.5 on 10 kinds of my real work. even if it wins, i'm still not letting it do the work.

why it matters

AI just got a pay scale. there's a boss rate and a worker rate, and i think most of us are paying the boss to do the worker's hours.

Anthropic's own launch page says start with Opus 5.5 for most coding and knowledge work, and keep Fable for the most ambitious open-ended stuff.

there are two cost stories and both are true. Anthropic says Fable on high matches Opus's best for roughly half the cost per task. Artificial Analysis measured Fable at about 2.6 times the cost per task at max effort. so the effort setting moves your bill.

what you get

  • the setup. Fable 5.2 in the boss seat and Opus 5.5 as the workers. two steps in Claude Code.
  • one prompt you can copy. it makes an AI judge two answers without knowing who wrote them, so you can test models on your own work.
  • the Fable reset catch, with dates. when to press the button so you don't waste it.

the real example: my run

i ran Fable 5.2 against Opus 5.5 on BuilderBench, my own benchmark. 10 kinds of my real work, marked blind, so i didn't know which one did which.

i made my bet before i saw a single score: Fable wins my bench, but barely. and i don't think it will matter much, because i'm not buying Fable to do the work.

the result is in the video at 10:35. one catch: Fable hit a usage limit mid-run, so the scores carry an asterisk.

the first thing i did with Fable was get it to plan this episode. it wrote me a take and i told it "that sounded like AI slop". AI can do a lot of this stuff, but i do write my own take.

in T3 Code i run the workers as separate threads, not subagents. about 12 hours in, the workload already feels lighter. that's a feeling, not a measurement.

try it

  1. open Claude Code and type /model fable. that puts Fable 5.2 in the main chat. that's the boss seat.
  2. set your subagents to Opus. in each subagent's file, set model: opus. those are the workers.
  3. start the boss on high effort. that's what i've been using and it's been working really well. medium is the default.
  4. give the boss one real job from this week. let it read the job, pick the worker and check the output.
  5. then test it blind. give two models the same brief. label the answers A and B so you can't tell which is which.
  6. paste the prompt below with both answers. pick the one you'd ship before you look.

copy this prompt

paste this into Claude or any AI, with your two answers under it:

Lenny’s AI Builders / LAB 0040
here are two answers to the same job, labelled A and B. you don't know which model wrote which and neither do i. for each one tell me: did it do the whole job, what's wrong with it, and would you ship it as it is. don't guess the model. then i'll pick the one i'd ship before you tell me yours.

my test in the episode ran on BuilderBench, my own tool. this prompt is the small version you can run today. i didn't run this exact prompt on camera.

you'll know it worked when

you picked A or B before you knew which model wrote it, and you can say why in one line. do that on a few kinds of work and you'll know which model gets which seat on your own tasks.

a bad result: you already know which model wrote which answer. then it isn't blind. swap the labels and go again.

limits and dates

  • one job is one data point. repeat it on a few kinds of work before you move your money.
  • effort changes the bill. the 2.6 times is Artificial Analysis at max effort. the half is Anthropic at high. compare at the setting you'd really use.
  • the price, as of 10 November 2026: Fable 5.2 is $8 in and $40 out per million tokens, down from $10 in and $50 out.
  • my bench has an asterisk. Fable hit a usage limit mid-run. and it's my bench on my jobs, not yours.
  • no score existed when i filmed. i made the bet first. the reveal went in about 30 hours later.

the rest of this LAB

  • Fable resets have a catch. Anthropic is handing out Fable resets on the top plans by how long you've subscribed. under 6 months gets 1. 6 to 12 months gets 2. over a year gets 3. they expire March 8, 2027. Pro and standard Team get one week to try Fable, up to half their weekly limit, plus one reset that runs out November 16, 2026. the catch: a reset only gives back what you spent on Fable. press it when your week is 99% Opus and you get basically nothing. so right after your weekly limit resets, use your Fable first, then press the Fable reset button. to find it, go to settings, then usage, then limit resets. once you press it, you can't undo it. my take: in LAB 0014 i counted 58 resets and called it a casino. now they've turned the casino into a loyalty program.
  • Grok Bot reset weekly limits for everyone. this came after Elon said the cost of serving Grok Bot was "unsustainably high relative to revenue" and that fixes for token efficiency are underway. so far i've noticed my standard prompts think much longer and come back much higher quality. that's what i've noticed, not a measurement, and it's too early to tell if the limits are fixed. my take: i'll believe it when i see my usage meter.
  • OpenAI's dot runs on several of your computers. you can connect more than one computer to the same dot and choose which one it uses for a task. it can build and test your app on a dedicated machine while you keep coding on your laptop. i've been doing this the hard way for months with two Mac minis. my take: your laptop stops being the bottleneck. if you've got an old Mac or PC in a cupboard, that's a worker now.
  • Meta's Muse Ember. Meta's smallest and most efficient model. it has a 1 million token context window and takes text, images, audio and video. Meta says it's nearly as smart as Muse Spark 1.3 at max for 17 times less, citing Artificial Analysis. that's Meta's claim. i haven't tested it. my take: Fable is the conductor and Ember is what the workers are about to become. so brain dump the tasks you'd never give an AI because the tokens felt wasteful.
  • the Jensen Huang, Nvidia and Megaspeed story. reported by the FT and the NYT. Nvidia says the Megaspeed claims were investigated and debunked. nothing was proven when i filmed.

sources

next step

watch LAB 0040 on YouTube - the bench starts at 10:35.

want the next one? get the next field notes. optional - these notes stay free to read.

want the next one?

free. unsubscribe any time, one click. privacy.