AI, biology, whatever I’m into.
Joined May 2025
- Tweets960
- Following110
- Followers111
- Likes3.3K
Pinned Tweet
Fable 5.5 and @threejs in a single HTML file!
Told ya Crouton was a Muse model :)
Crouton Alpha is no where near Pulse Alpha...
Although it one shotted it and did very little browser verification on its output, Pulse spent most of the time doing browser verification and perfecting its output.
Pulse seems a bit reckless like DeepSeek while Crouton seems pretty chill and not that good in 3D as if it were a Meta Muse model.
9
Crouton Alpha is no where near Pulse Alpha...
Although it one shotted it and did very little browser verification on its output, Pulse spent most of the time doing browser verification and perfecting its output.
Pulse seems a bit reckless like DeepSeek while Crouton seems pretty chill and not that good in 3D as if it were a Meta Muse model.
gpt plus is still scamming us, 13 minutes of codex ate my entire 5h limit
GPT 6.1 Sol xhigh, fast mode, 1 main + 2 subagents
5h window: 0% to 100% in 13.2 min
weekly: 17% to 33%
the thinking only cost $0.28. the real burn was codex re-reading ~170k tokens of context on every step, in 3 agents at once
if you're on plus:
- turn fast mode off for long tasks, it doubles every token
- don't fork subagents with full history, mine started at 113k-141k tokens instead of ~20k
GPT Plus is the biggest scam of all time🤡
Ran GPT 6.1 Terra max in codex, same prompt, normal vs fast
normal: 15m53s, 81.5 tok/s, $0.71
fast: 17m02s, 79.0 tok/s, $1.81
fast costs 2x, got slower and same tok/s
@thsottiaux you should probably retire the Plus plan now, we go thru all the bs, 5h limit, no doubled usage, no dots, no GPT Pro models
I can imagine the unemployment rate in India people will do anything to make a point and prove you wrong 😂
As I was reading this reply.... openrouter.ai/anthropic/clau…
sometimes, its ok to accept that you are humbled and move on.
What are you kind of high are on when you are saying you saw Sonnet at 150 Tokens/second and Haiku on 250 Tokens/second.
Cleary your mod or whatever that is displaying the speed is not correct, fix that to see the correct numbers.
2
108
GPT Plus is the biggest scam of all time🤡
Ran GPT 6.1 Terra max in codex, same prompt, normal vs fast
normal: 15m53s, 81.5 tok/s, $0.71
fast: 17m02s, 79.0 tok/s, $1.81
fast costs 2x, got slower and same tok/s
@thsottiaux you should probably retire the Plus plan now, we go thru all the bs, 5h limit, no doubled usage, no dots, no GPT Pro models
GPT 6.1 Terra max vs GPT 6.1 Terra max fast in codex
normal: 15m53s, 81.5 tok/s, $0.71
fast: 17m02s, 79.0 tok/s, $1.81
same prompt, fast costs 2x and came out slower
fast mode on plus is doing nothing bro. either it's not working or @thsottiaux is lying to plus users 🤡
(api-equivalent cost, i'm on plus)
GPT 6.1 Terra max vs GPT 6.1 Terra max fast in codex
normal: 15m53s, 81.5 tok/s, $0.71
fast: 17m02s, 79.0 tok/s, $1.81
same prompt, fast costs 2x and came out slower
fast mode on plus is doing nothing bro. either it's not working or @thsottiaux is lying to plus users 🤡
(api-equivalent cost, i'm on plus)
GPT 6.1 Terra max vs GPT 5.6 Terra max in Codex
5.6: ~22mins, 6.56M tokens, ~$2.42
6.1: ~16mins, 5.79M tokens, ~$0.71
I am genuinely thinking of re-running 6.1 because this output seems unnatural for a model this good.
9
64
14,070
obsidian 3 isn't fable 5.5, my bet is it's an opus 5.6 checkpoint
it showed up hidden in cognition's frontiercode 2.0 data.json, and now it's gone from the live file 🫣
high: 46.8%
xhigh: 45.9%
max: 45.5%
opus 5.5 max: 45.5%
it only runs in the claude code harness, so it's anthropic
it lands right next to opus 5.5, not a big jump like a new fable would be
and it gets worse the more you let it think, feels like an early checkpoint
could be wrong, dyor tho
sources:
found by chetaslua: xitter.gq/chetaslua/status/21153…
cognition.com/data/frontier-…
🚨 Obsidian 3 is hidden model in FrontierCode 2.0
This is the highest scoring model : what can it be < a Fable 5.5 or something else>
Harness: it runs on claude-code, which only the six Claude models use
Price level: it costs about 3.6× what Opus 5.5 costs at the same effort.
1
65
