Also, the frontier labs are probably always internally using a model more advanced than what has been publicly released.

▲ user43928 3 hours ago | parent | next [-]

From the model card:

  Model                AECI   95% CI (global)
  Claude Opus 5.5    174.78   170.33-180.29
  Claude Mythos 5.2  177.12   172.42-182.94
  Internal model     178.78   173.75-185.18

Considering the IPO is planned in a week, I doubt they are holding much back.

▲

jeremyjh 2 hours ago | parent [-]

According to their S-1:

In H1 2026, cost of revenue is $2.3B. That is primarily inference compute, but also includes cost of training models (but not research).

In the same period, R&D was $9.4B. That is primarily spent on compute. That is what they are "holding back" already, in full public view.

This is not a company that is optimizing for OpEx or end user revenue. The only thing relevant to their plans is achieving AGI/ASI first, and controlling enough compute to leverage it to earn the majority of the current economy.

▲

sfink 29 minutes ago | parent [-]

"Achieving AGI" strikes me as a mostly meaningless phrase. I'm skeptical that it is a boolean event. They're already smarter in some ways, their impact is large and gradually rising, the set of specific tasks that they aren't capable of is gradually shrinking. If AGI means anything useful, it seems like we'll only see it in hindsight and have to argue over when "it" happened.

Likewise the singularity. While recursive self-improvement is theoretically trivial (just tweak a random subset of weights like a monkey with a typewriter, eventually you'll get something smarter), the resulting exponential improvement is still very hypothetical and feels unlikely. Unless you're ok with an exponent of 1.000000000000000001.

	▲	jeremyjh 7 minutes ago \| parent [-]
		These are not my stated ambitions, they are the ambitions that Anthropic and Dario have said they are pursuing. Whether or not they are realistic is not relevant to this thread. The credibility of the plans are relevant to the S-1 and anyone who purchases stock from them, but for the purposes of this thread, it is only important that Anthropic believes it has credible chances of achieving this goal, and in particular that it has a competitive advantage, could do so first, and that being first could have a very large upside.

▲ vlyan 3 hours ago | parent | prev [-]

because they hate money? google is the only frontier lab that isn't hundreds of billions in the red.

	▲	jeremyjh 3 hours ago \| parent \| next [-]
		Because they do months of testing. Also, if they have models that are really good at AI research, I’d not expect full versions to ever be released, precisely because of their love of money.
	▲	thomasmurph 3 hours ago \| parent \| prev \| next [-]
		Inference on the research models they have is likely more expensive than the released ones.
	▲	93po 2 hours ago \| parent \| prev [-]
		My assumption is that they're overly expensive to run at scale and/or don't have the same safeguards that public models have and possibly get better results because of that. If a model is 100x more expensive to run but 4x better results, of course you'd use that internally to make your product even better - it's compounding returns. But people aren't going to pay 100x their current bill for 4x results.