GLM 5.3 has roughly the same tokens/outcome efficiency as Opus 5, so the tokens/sec rating pretty much translates 1:1 to "actual outcome speed".

Well: 1. EXL3 4bpw isn't "retarded". 2. For now, the hardware prices continue to climb, so it's the opposite of "negative ROI" if you sell sometime before they fall. 3. In terms of TPS, the 50-60tok/s range has been pretty standard until recently. Example:

Today's PHP vulnerabilities are the outcome of decades of human slop.

As low as $24k if you bought Asus GX10s for their launch price of $3k each late last year. Even as recently as August, GX10s were $4k each, putting you at $32k for 8 for them. So "what did I miss" = All of the decent prices before the dramatic price swing lol.

But if you tell it "uh it's definitely right here and it exists" then it can find it lol.

Straight up tagged in Pi and it can't find it lol.

If you trust their WER mix for your use-case, then this is the rare situation where Europe is the open-weights lead lol.

This also feels like an itty bit of bias in favor of "renting vehicles rather than owning them" lol.

Yeah AFAIK from their methodology page, it doesn't seem like it's scored well. Not a single human involved in evals that are meant to express how a model comes across to a human... lol. artificialanalysis.ai/method…

Interesting: GLM 5.3 Flash and DeepSeek 4.1 Flash 1017 live toward opposite poles of AA's hallucination score.

By contrast, this staggered road is very clearly taking more liberties with allowing each stretch of road be more granular and individual, more in service of a given high-level product outcome rather than prioritizing the "taxonomy purity" of the systematizing mind.

This right here is what happens when you have a bottom-up approach of defining what a "road" looks like as an atomic unit in code, and then copy/pasting that atomic unit indefinitely. It aligns with DRY ("Don't Repeat Yourself") coding principles: