I was wrong
Sol looks efficient on Artificial Analysis, but in real-world use it uses more tokens than Opus
When I just had aggregate numbers for the whole model, I thought this was because people tended to use Sol at higher reasoning efforts than Opus
But it turns out the split is pretty similar
About half of use is done at high reasoning for both
But despite all the benchmarks, Opus ends up significantly more efficient on high reasoning (the most common)
These are still different tasks
So usage patterns between Opus and Sol are different, but it's not because of different reasoning efforts
Show more
The best open model on GDPx is DeepSeek 4.1 Flash, at less than half of Opus 5.5
It's closer to a third if you don't count the domains where Opus refuses
Open models haven't been this far behind in a while
Show more
The smarter a model is, the more subagents it spawns
I noticed this before but the correlation is stronger than I expected
I held off on switching for a while bc there were a few things Codex had that T3 Code didn't.
What I failed to account for is the things T3 Code has that Codex doesn't.
Never looked back
How often users interrupt the model is a pretty interesting benchmark.
Haven't seen this anywhere except T3 Code data from Slopalytics.
Why is my feed so suddenly full of political stuff?
Where are my graphs and charts?
I guess I'll have to make up for it with a hype reel
Google is the only top AI lab without an xhigh or max reasoning setting
ChatGPT writes 6.7x more than Meta Muse to answer the same question
Muse's personality is underrated
This is what it looks like with subscription subsidization turned on
Gemini is the funniest model
4o was truly a special model
I just told bot literally "go do this for me"
this was the result
Ranking subscriptions by how worth it they are
DeepSeek V4.1 Flash is 4.4× faster than DeepSeek V4.1 Flash
I'm looking for a reliable source of data on subscription subsidization across lots of different models so that I can add this data to priors.
Anybody got a good suggestion?
Cost per task metrics always fail to account for subscription subsidization.
You can get way more bang for your buck on the $200 plans than paying API rates.
If they cut the API price of Opus in half tomorrow, but didn't affect your subscription usage, the "value" of that subscription would be cut in half without you losing anything
All of these numbers are just inflated
Show more
ok, how'd I do?
ㅤ DEVDAY-RECEIPT
---------------------------
Astra 6.1 ....................… 75%
$500 plan …................ 90%
GPT Live 2 …................ 30%
GPT-Image 3 ...........…. 30%
GPT-6o …...................... 20%
Bel announcement … 40%
"o" ...............................… 99%
Ultrafast ....................… 20%
OpenAI hardware …... 50%
Bel release ................... 0.1%
Hodge solved .............. 50%
Astra Major/Minor ..... 6.7%
"one more thing" .....… 80%
TOTAL ....................... HYPE!!
Show more
Mythos 5.1!
wen Fable 5.5 tho?
We’re expanding our Cyber Verification Program to give security professionals broader access to our most capable models.
Through this program, verified security professionals can access Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 with safeguards designed for defensive work.
We’re also opening up new tiers to allow for authorized offensive work, like penetration testing and red-teaming.
Show more
Thanks Tibo!
Day 2.2/
OK this one is smaller, but we removed some friction for when building on the API. Happy building.
So we turn everything to max, pay full API prices, and then call the model overpriced?
