software engineer turned vibe coder. always on a side quest
Joined July 2026
- Tweets202
- Following245
- Followers275
- Likes925
when you press send on your perfect prompt except you spelled it "architechture" and now the model thinks you're not a serious person
Haiku 5.5 vs. Luna 6 on xhigh
Prompt: A 10-second flat-shaded animation of a pastel kitchen with a mint-green toaster. The toaster winds up like a cannon and launches a slice of toast. The camera whip-pans after it, the toast spins with rainbow trails, cabinets jiggle like gelatin, fridge magnets fly, crumbs storm, and butter bursts into golden sparkles. It lands in slow-mo and sends a ripple across the whole kitchen. Hard cut to silence: the toast rests neatly on a plate, and the exhausted toaster puffs one smoke ring and goes "boop." Build in a HTML then return an mp4 in portrait mode.
This is one of my biggest issues with OpenAI models and one of my favorite things about the latest Opus model (and Fable).
They DO before responding. It may seem small but it drives me crazy. Of course you can’t know, you haven’t checked yet…
This would be a huge win if you guys could fix that @thsottiaux @OpenAIDevs
God @OpenAI what'd you put into this thing..
Did you RL it to love Fatheads or something?
Introducing Gerald, our most agentic agent yet.
You: book me a flight to Denver
Gerald: LOVE this. What does Denver mean to you?
He's bought a horse and proposed to Denver. Denver wants to take it offline.
Made with Claude Opus 5.5, Sonnet 5.5 and ElevenLabs v4
Probably Nick retweeted
I was spending $1,200/mo on Codex plans -- six $200 subscriptions
after Opus 5.5 one Claude $200 subscription handles the same amount of tokens
plus MORE, due to how insanely fast Opus 5.5 really is
so yes, this is fairly accurate
Pretty frustrating that this didn’t happen. I ran some Astra ultra agents to get through my usage assuming it would reset at 10am PST only to return to no usage & a large portion of my usage credits gone.
Rough comms week for OpenAI
Everyone is running wild with the Fable 5.5 Tibo test.
Fact check:
1. Tibo was known for resetting as far back as April 2026 xitter.cf/thsottiaux/status/2048…
2. Opus 5.5 has the same June 2026 knowledge cutoff as Fable 5.1 so we have little proof Fable 5.5 would have a later one
Gotta love how quickly misinformation spreads on here..
My test results do suggest it could be routing: xitter.cf/nickfromlater/status/2…
I stand corrected. I stress-tested the Fable 5.1 “Tibo” test:
API Fable 5.1 went:
0/105 on “Tibo the reset guy”
99/105 once I added “Codex,” so it definitely knows who he is.
Claude web Fable 5.1 answers it consistently without “Codex” and I even tried adding the published web system prompt via API but that didn’t change it.
Maybe it really is routing. The visual results I’m seeing seem to back that up too. Exciting times.
I stand corrected. I stress-tested the Fable 5.1 “Tibo” test:
API Fable 5.1 went:
0/105 on “Tibo the reset guy”
99/105 once I added “Codex,” so it definitely knows who he is.
Claude web Fable 5.1 answers it consistently without “Codex” and I even tried adding the published web system prompt via API but that didn’t change it.
Maybe it really is routing. The visual results I’m seeing seem to back that up too. Exciting times.
With that said, the results I'm seeing on my feed are impressive, even if the majority (if not all) of them are Fable 5.1
While I’m impressed by the results I’m seeing, I’m not really sure why “Who is Tibo?” or “Who is the codex reset guy?” is considered a reliable cutoff test
Tibo was resetting codex limits way before Fable 5.1’s June 2026 knowledge cutoff.
Here’s a post from April 2026… xitter.cf/thsottiaux/status/2048…
🚨 Fable 5.5 is auto routing on web , this is the screenshot it edited for x without even prompted
he knows tibo check claude.ai see if you are getting routed or not
