@gabepereyra@harvey So my naive question would be — from the customers perspective, what’s the difference between using Harvey with frontier models + Harvey intelligence layer vs custom models? Both setups can take advantage of customer data to deliver better outcomes, so is it purely a cost thing?
“Intelligence does not exempt a system from the realities of the physical world!”
There is no equivalent of bash in the physical world. Humanity is just fine.
@katelyn_lesse do you think a similar price squeeze will happen on data center CPUs then? seems unlikely, but there's a version of this where the actual bottleneck in scaling your agentic AI compute stack is CPU not GPU
@siddharthvader_ Chip design and formal are particularly interesting. Iteration loops and time to verify are quite long, so these tasks tend to be both extremely token intensive and long horizon.
@matanSF Curious what you mean by post training on tool use being not exclusively advantageous. In what situations is it advantageous, and in what situations is it not?
This is Claude Sonnet 4.6: our most capable Sonnet model yet.
It’s a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design.
It also features a 1M token context window in beta.