Post

Almost certainly GLM. Maybe something like GLM-5.3V-Turbo. I think it's strong in coding but otherwise not on Sol or Fable level. Thinks a lot, makes plenty of errors in thinking (which is shown, so open model), doesn't seem too good in math, vision not that good, sees imagined typos in PDFs.

Haven't we been here before, regarding secret models turning out to be a nothing burger?

what kind of psychopath did this? idk if whoever ran these can’t get basic details right, i doubt the benchies are right either

Free access makes this an interesting cost–quality datapoint, but a 1M context claim won’t tell us much about coding or agentic reliability by itself. I’d want task-level tests, answer rates, and repeatability before drawing conclusions from a stealth endpoint.

Usually those free stealth model quickly get severely overloaded and it is hard to get a request through. At least that was my experience in the past.

they saw the Sol discount and said hold my beer

that’s the dominant gossip

No data usage either. It'll be really funny if it's Grok 4.7 or something

I still really like GLM 5.2, so if it is GLM that'd be very cool.

argument against z.AI so… SSI??

If anyone's wondering, openrouter doesn't need a payment card to sign up. Thought it would, but it's fine without one.

It's available free in "command code" site and harness now too which I'm currently evaluating:

i'm gonna say it's from Cohere because I like them