Stefano Caioni

Group Product Manager @freelancer
Still write code, now with AI. Shipped LLM features to millions.
Building agents to shrink my week.

[Stefano Caioni @stefano_caioni]

1/

I had Claude read X's open-sourced “For You” code.

What actually decides reach:

A reply, quote or DM share carries 10x the weight of a like. A copy-link share, 40x. Bookmarks aren't in the score at all.

If you write for likes, you're chasing the smallest number on the list.

[Stefano Caioni @stefano_caioni]

Reading the plan carefully matters more than reading the code.

Bad code takes a minute to fix. A bad plan can cost you weeks.

[Stefano Caioni @stefano_caioni]

@hectorserrranoo @whop @czoob3 Congrats!

[Stefano Caioni @stefano_caioni]

@luisnhenrich Agreed. Cost per task is just another eval score, so it needs the same reruns on your own tasks.

[Stefano Caioni @stefano_caioni]

Fable 5.2 shipped yesterday.
The 1,000-agent demos are impressive. But if you build products with agents, here's what actually changes.

1. Long-running work is getting real.

Anthropic says 1,000 Fable agents formalised an existing proof of Fermat's Last Theorem in 30 hours. With the same setup, Opus 5.5 had done 26% at 18 hours and had barely moved by 30.

It's one controlled experiment, so no promise for production.
Still a serious jump.

2. Cost per completed task beats cost per token.

At $8 input / $40 output per million tokens, Fable costs twice as much as Opus 5.5. But Anthropic reports that on one coding eval it matched Opus's best result at roughly half the cost per task.

If a cheaper model needs three retries and a human fix, was it really cheaper?

3. Don't make it your default.

Anthropic calls Opus 5.5 its daily driver. I'd start there and escalate when a task keeps failing, needs deep reasoning or takes too much human supervision.

4. Orchestration is becoming a product decision.

One Anthropic engineer had Fable coordinate over 1,000 subagents for eight days, from an empty folder to a prototype operating system that boots on a laptop. Fable can also hand simpler work to smaller, cheaper models.

So deciding which model checks the work is now as much a product call as picking the model.

5. Evals need to look more like production.

Anthropic says the new models voice suspicion that they're being evaluated more often than its recent models. That doesn't prove they can spot tests, but it complicates how we read eval scores.

Test with real user tasks. Run them repeatedly. Track failures, recoveries, and cost per successful outcome.

What's one task you still won't let an agent do unattended?

[Anthropic @AnthropicAI]

Claude Fable 5.2 and Claude Mythos 5.2 are available now:

[Stefano Caioni @stefano_caioni]

@shafu0x Wasn’t it growth hacking?

[Stefano Caioni @stefano_caioni]

@nicomoel Willing to listen and willing to ask the right way

[Stefano Caioni @stefano_caioni]

@seeksteve Honestly labelled slop is the best kind of slop. Does the dog have an album?

[Stefano Caioni @stefano_caioni]

@oderajoseph Simple but brilliant. Well I will make my own haha

[Stefano Caioni @stefano_caioni]

@icanvardar Surprised no one has built a directory for dead projects yet.

[Stefano Caioni @stefano_caioni]

@PrajwalTomar_ Step 5 is the clever one. Put your brand name inside the stat, and ChatGPT has to say your name every time it quotes it

[Stefano Caioni @stefano_caioni]

@hakimieiqbal Winning ads already wear out in weeks. Now anyone can clone yours shot by shot, so call it days

[Stefano Caioni @stefano_caioni]

@DataconomyMedia 81% of the visits but only 60% of the clicks out to websites. Looks like the smaller chatbots link out a lot more

[Stefano Caioni @stefano_caioni]

@TheSaaSCraftman With ads already in ChatGPT, this is Search Console and Google Ads 2.0

[Stefano Caioni @stefano_caioni]

@mardehaym I'd drop time per approval though. Good reviewers are fast because their agents and rules already caught the easy/recurring stuff. Not every output needs a human either.

[Stefano Caioni @stefano_caioni]

@cyberandy @wordliftit The outside half is the complicated one. You can fix your own agents' memory, but fixing what ChatGPT knows or remembers about you is way harder

[Stefano Caioni @stefano_caioni]

@leonabboud As predicted. That video on his profile and the interview Lenny’s podcast were gold

[Stefano Caioni @stefano_caioni]

This is wild. Beautiful. And wild.

[Mia @MiaAI_lab]

I asked Claude Fable 5.2 to create 100 HTML files

Rules were simple:
- look stunning
- zero repetitive designs
- go full creative mode

What can I say? Anthropic has done it YET AGAIN.

Just another mind-blowing model, nothing less. Fable 5.2 is the BEST MODEL I HAVE EVER TESTED on frontend web design‼️

The previous best result on this experiment was Opus 5.5, but Fable 5.2 is a noticeable step up.

Many files include playable games, fun experiments, and just insanely high-quality HTMLs and designs.

There is ONE big problem though: it costs 2x as much as Opus 5.5! This experiment includes only 76 files, because this prompt alone used ALL of my Fable usage, including a FULL RESET!

Meaning it burned TWO weekly Fable usage limits!

However, as expected, ZERO BROKEN FILES.

There are TOO MANY INSANELY GOOD files to list them all here. YOU MUST go over EVERY SINGLE FILE.

See the results for yourself (incl. prompts):
https://t.co/nKVBRUpWWK

All other experiments here:
https://t.co/HVhecHtyzd

Some great examples:

Volcán Chico Hot Sauce
https://t.co/5Jtvj4RU1G

Koi Pond
https://t.co/xOCJaKslVY

Descent - A Vector Lander
https://t.co/qGpPhgohPJ

Landscapes in One Pen - this is amazing.
https://t.co/LYCF3VD34E

The Wheel - Throw a Pot
https://t.co/iusJzXCN7q

Tangram Menagerie
https://t.co/vlPBc95V3g

The Night Typist
https://t.co/87UX59r6u1

Lava Lamp, 1971
https://t.co/9d4EG5oWHm

Jelly Alphabet
https://t.co/GMFjQJM4Gn

Ariadne's Thread - just wow!
https://t.co/kpydBBETx1

Cross-Stitch Studio
https://t.co/L0Ap9tr4Yd

Pinscreen
https://t.co/S2nPbhVYfd

And so on... just open all the files!!!

[Stefano Caioni @stefano_caioni]

@georgeorch Yea I find it even weaker on actual images or illustrations etc

[Stefano Caioni @stefano_caioni]

@georgeorch Aesthetic judgment. Still feels they give pretty generic advice

[Stefano Caioni @stefano_caioni]

@pmitu When AI governments?

[Stefano Caioni @stefano_caioni]

@PawelHuryn I wonder how much better the results would be if you gave the agent some product context first. Who the users are, when they hit their aha moment, and where they experience the core product value. Then let it check the code