Hacker News Reader

Comment by minimaxir | Hacker News Reader

No pricing, but given that OpenAI's GPT-4 doubles the cost-per-token if you go from 8k to a 32k context window, I suspect the pricing here will be 2-4x from the base Claude model which is 9k: https://cdn2.assets-servd.host/anthropic-website/production/...

Although with flash attention, who knows if marginal cost scales that consistently.

View on HN

Pricing is the same as the base model.

Confirmation here:

https://twitter.com/AnthropicAI/status/1656743460769259521?s...

Huh. Well that changes things.

Only for the duration of the beta

Source?

the actual tweet you linked.

It doesn’t say exclusively for the beta period

With an extremely literal reading you are correct, but there was clearly an implication.

<4x would be quite optimistic, at ~11x the tokens the amount of compute/memory required would be n^10 (even with the lower starting point of flash attention) so unless they are already have excessive margins it wouldn't make much sense to go that low.

I was assuming they used a different architecture to get the increase instead of just letting it eat hardware that way. Especially with the speed numbers in the post.

https://cdn2.assets-servd.host/anthropic-website/production/...

Those are the same SKUs I linked.

The new model are a different model identified that's not listed in the pricing doc, although it sounds like the intent may be to replace the base from looking at the API docs: https://console.anthropic.com/docs/api/reference#-v1-complet...

Reply on news.ycombinator.com