Repo logs and caching price
swe-bench is pretty good at sharing the eval logs. Do you do this anywhere?
Also, are your token prices for the chinese models correct? (glm, minimax) They have very good cache read savings that I don't think you're taking advantage of. Look on their website or specific providers on openrouter (eg, deepinfra)
Ideally you'd always be going to the company website as api provider, tbh, as otherwise you can't be reasonably assured that the deployment was correctly done, a big problem with chinese models and 3rd party providers.
I think you guys have some good ideas, but the low investment in maintaining it and lack of consensus based task chosing severely undermines your credibililty and probably doesn't do favors for your Nebius brand. my suggestion would be to either spend more (get grants), or fold it into swe-bench. Required though would be getting a credible panel to manage task choices. Without the latter, these benches are quite useless
Hi, thanks for the feedback. At the moment, we run open-source models via OpenRouter without prompt caching because most providers on OpenRouter do not expose caching yet. However, we plan to add direct evaluation through official providers soon, which will allow us to enable caching wherever it is supported. Thanks for highlighting this.
OpenRouter often exposes cache pricing if you pin it to a specific provider.