Adam Holter

@AdamHoltererer • 24,767 subscribers

taking dominion of AI sponsored by @AmplifyPartners

Shorts

Fable 5.2 is a huge jump in TakingYourJobBench

104,056 views

$Opus 5.5 fails twice as many T3 turns on max as on high My first guess would be that it's because turns on max, take longer, but no: max fails more in every turn-length bucket, and most of the extra failures happen in the first 30 seconds. Sol shows the same pattern, smaller (1.4x) So why is this? ¯\_(ツ)_/¯$

Opus 5.5 fails twice as many T3 turns on max as on high My first guess would be that it's because turns on max, take longer, but no: max fails more in every turn-length bucket, and most of the extra failures happen in the first 30 seconds. Sol shows the same pattern, smaller (1.4x) So why is this? ¯\_(ツ)_/¯

117,107 views

Oh, thaaat's why they use a logarithmic scale!

900,634 views

Gemini 4 Hydrogen is the most middle-of-the-pack model ever

13,112 views

OpenAI has reset Codex limits 11 times this month

244,267 views

Google's plans have terrible subsidization Even Opencode subsidizes Argon more than Google's own plans do

113,886 views

The AI market right now

114,193 views

yolo mode ftw

59,414 views

Artificial Analysis' cost numbers for Gemini are inflated right now Google doubles Carbon and Argon's price on Dec 1, and AA's charts already use that However, the rates that users are paying don't reflect either of those numbers

18,695 views

...if you don't consider subsidization

20,958 views

Hot take: Tokens per second is a completely meaningless metric

303,047 views

Google is doubling the price of Gemini 4 Argon and Gemini 3.8 Flash Really weird since they'll both be obsolete by then anyway.

100,126 views

How small is o1-level intelligence now? 3.7B or 25B, depending how you measure. It's impossible to measure raw intelligence, so you have to use some kind of benchmark index.

26,088 views

I'm using T3 Code more than I ever used Claude or Codex

55,970 views

ChatGPT writes 6.7x more than Meta Muse to answer the same question Muse's personality is underrated

93,304 views

Carbon's cheaper at API prices Nobody pays API prices

12,847 views

Claude has resets too They're just a lot less frequent But they don't put your reset date a week out tho. It's strictly a bonus, which is nice.

74,270 views

Mistral may have trash models, but Vibe subsidizes GLM 5.3 way better than does

56,052 views

Cursor's new $200 plan includes zero Opus 5.5; Claude's own includes about $13k of it. Existing Cursor customers keep their old pool for now. New Ultra users have to pay for Claude on top at API rates.

48,884 views

On Artificial Analysis, GPT-6.1 Sol uses about a third of Opus 5.5's output tokens at the same effort, and half if you account for score. In real T3 Code use, Sol spends 41% more output tokens per turn than Opus.

60,398 views

Videos

LIVE

1.2k

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Streaming Now

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

HD live stream

Exclusive private shows

1.2k viewers online

Current Status

Live

Private Show

Join now for exclusive access

Free preview available • Premium content

Live Cam

Adam Holter

Shorts

Fable 5.2 is a huge jump in TakingYourJobBench

Oh, thaaat's why they use a logarithmic scale!

Gemini 4 Hydrogen is the most middle-of-the-pack model ever

OpenAI has reset Codex limits 11 times this month

Google's plans have terrible subsidization Even Opencode subsidizes Argon more than Google's own plans do

The AI market right now

yolo mode ftw

Artificial Analysis' cost numbers for Gemini are inflated right now Google doubles Carbon and Argon's price on Dec 1, and AA's charts already use that However, the rates that users are paying don't reflect either of those numbers

...if you don't consider subsidization

Hot take: Tokens per second is a completely meaningless metric

Google is doubling the price of Gemini 4 Argon and Gemini 3.8 Flash Really weird since they'll both be obsolete by then anyway.

How small is o1-level intelligence now? 3.7B or 25B, depending how you measure. It's impossible to measure raw intelligence, so you have to use some kind of benchmark index.

I'm using T3 Code more than I ever used Claude or Codex

ChatGPT writes 6.7x more than Meta Muse to answer the same question Muse's personality is underrated

Carbon's cheaper at API prices Nobody pays API prices

Claude has resets too They're just a lot less frequent But they don't put your reset date a week out tho. It's strictly a bonus, which is nice.

Mistral may have trash models, but Vibe subsidizes GLM 5.3 way better than does

Cursor's new $200 plan includes zero Opus 5.5; Claude's own includes about $13k of it. Existing Cursor customers keep their old pool for now. New Ultra users have to pay for Claude on top at API rates.

On Artificial Analysis, GPT-6.1 Sol uses about a third of Opus 5.5's output tokens at the same effort, and half if you account for score. In real T3 Code use, Sol spends 41% more output tokens per turn than Opus.

Videos

Watch Anya Live

Looks almost competitive until you consider subsidization

Ranking inference providers six ways

Astra fails to clone this and just makes a 3D voxel game instead of a 2D rendered game built with 3D

Hot take: Tokens per second is a completely meaningless metric

How small is o1-level intelligence now? 3.7B or 25B, depending how you measure. It's impossible to measure raw intelligence, so you have to use some kind of benchmark index.

Carbon is the sixth element. Here's all the others: Gemini 4 Carbon High: Periodic Table Song Test

Cost is a bad thing It should be down oh no... 💀💸

Grok 4.7 thinks its company is called XDI

The problem with AI is that it doesn't believe in itself

Terra is the best sub-agent model ever By far the best speed vs performance trade-off for your main agent to delegate to

OpenAI has historically had the best model on AA But the tables have turned...

DeepSeek V4.1 Flash breaks tool calls 45x as often as DeepSeek V4.1 Flash Providers can matter a lot

Big update to priors: resets are now included in the subsidization data This brings OpenAI back to the Pareto Frontier! Thanks, Tibo !

Jev versus GPT 5.6 Luna (low) in chess Jev is white, Luna is black

OpenAI, please let us customize our default slider. I want to be able to set it up with: • Luna Max • Terra Light • Sol Light • Astra Low • Astra Medium • Astra xhigh

You can mod Claude Code now, so I did the obvious

Astra's first pass at cloning it

Astra tried to clone the visual style, but the gameplay is way off