WeirdML v2 benchmark finds Claude Sonnet 5 yields modest accuracy gains but major cost improvements over Sonnet 4.6

incredible

Sonnet-5 is lagging the frontier by 7 months

Nothingburger-5 scores on WeirdML

Also, it's presumably like half the size of GLM so, very good work on density too! It must be as cheap to serve as DSV4-Flash, but it'll have more use even at 40 times the price. Winning!

Nothingburger-5 is the new Llamao-4

@scaling01 Impressive results for DSV4-Flash scaled model!

@scaling01 I really feel everyone is pretending every model is supposed to be for coding.

I dont understand sonnet 5 to be for coding at all. Its for writing and general workflows

@scaling01 So is sonnet 5 just the most expensive email writer? I’ve noticed it’s good at doing some web fetches I suppose

@scaling01 nah the real fight is deciding which model is lying to you best

but frontier wars are fun to watch

@midsusnight I’d suggest you read the more detailed post Lisan made out in his telegram channel SCALINGCALLS cause from my understanding you’re getting him all wrong

@Greg_TheBuilder I’m sure you’re getting his view confusing, thought Same as you too when I saw the post here on x, check out the updated post on it Lisan made out on his telegram channel SCALINGCALLS I’m very sure you’ll understand him better after reading the post

@essenciverse I’m sure you’re getting his view confusing, thought Same as you too when I saw the post here on x, check out the updated post on it Lisan made out on his telegram channel SCALINGCALLS I’m very sure you’ll understand him better after reading the post

@essenciverse @scaling01 then get it out of claude code if it sucks for coding and finetune a good coding model

@teortaxesTex 把挤牙膏当成了顶级商战真是绝了