Qwen 3.5-35B already runs on a secondhand RTX 4090 and "matches" Claude Sonnet 4.5 on (some) benchmarks at $0.10 vs $3.00/mn tokens.

Did you comment on the wrong article? It's not about Qwen or Sonnet and no rtx 4090 is mentioned.

Not really but fair point.. was thinking more about the commoditization angle since I was just reading up on this, like how every generation of consumer GPU that can run near-frontier models locally makes closed API pricing harder to defend

The GB10 workstations are not that competitive in price with M5 MacBooks and the Mac Studio unless you really need CUDA.

I hope there will be an M5 Ultra at some point, but the geometry of the M5 Pro and Max don’t make it obvious (unless there is an wide interconnect on the other shore of the chip).