Post

For the Claude comparison, are you measuring cost per successful result on a fixed workload, with retries and latency included? The $3/$15 rate can look cheap until instability and retries enter the denominator.

no, I realize the list price of Kimi K3 is around half the cost per request on average compared with Opus 5 but on the €18 team subscription, you get way more with a month of daily usage than the list price - from what I've seen, it's a lot more than you'd get for €18 at list price, is it not?

Agree. My point is that token pricing is a poor proxy for outcome quality, so list prices are too. We benchmark actual task outcomes (llmbench.kapualabs.com), and the differences are huge. You can easily pay 100× more and still get a worse result.