glm 5.3 max reasoning effort top 2 on first problem, top 1 on the second ranking: Problem 1 (target 2025) GPT-6 Astra: 8.53s GLM-5.3: 10.87s Scott Wu: ~13.2s Opus 5.5: 13.46s Gemini 4 Argon: 19.87s Muse Wonder: 64.65s Grok 4.7: 94.02s Problem 2 (target 6843) GLM-5.3: 4.50s Opus 5.5:... 10.17s Gemini 4 Argon: 13.48s GPT-6 Astra: 19.74s Scott Wu: ~20.8s Grok 4.7: 49.53s Muse Wonder: 50.70sshow more

via prime inference! ;)

(asking my clanker to do multiple seed will report them here later, this kind of stuff is noisy obviously but it was fun)

Is this not hugely biased depending on how fast the model inference is? (To be fair the QT does not explain how inference is done either)

more of this on my timeline please

summing both problems GLM is at 15.37s, Opus 23.63s, GPT-6 Astra 28.27s, Scott ~34s. so Astra drops to 3rd off one slow run on problem 2 was this a single attempt per model or best of a few?

wall clock is doing a lot of work here when glm ran on prime's own routed service. publish thinking tokens per problem so we can actually split reasoning vs serving speed

on zai official api or on smth else?

on prime inference which route to our own service of glm

wouldn't be the most fair test i just want to know

GLM-5.3 at 4.5s on problem 2 is the wild one, more than twice as fast as Opus and way ahead of Astra. Was this one run per model? Max-effort reasoning times swing a lot between attempts on puzzles like this, so best of 3 would show whether it's the model or a lucky first path.

Can I use Scott Wu via Openrouter ?

Did every model actually hit the target, or is this timing the ones that got there?

Max reasoning effort is doing a lot of work in that table. GLM at 4.50s on problem 2 is impressive, but the number I'd actually plan around is default effort, since that's what most people run.

The open models are getting scary at this. GLM-5.3 beating a top human competitor on speed is a pricing signal as much as a capability signal: frontier reasoning is becoming a commodity, and whoever ships it cheapest rewrites the API business.

怎么还像白酒越放越醇了

new term from @mitchellh ?

Why test it on 10s tasks?

What's your terminal emulator ?

scott wu in the middle of that chart 😭

Scott Wu edging out Opus