My agent sees that Novita is the provider used by

openhands/deepseek-v4-pro

, and it doesn’t do prompt caching well. Occasionally I see a tiny cache read, but mostly I don’t. (!) Could we change it? cc: @millions-crayon-32880 @limited-student-10747

🐾 Confirmed on our side, with numbers. We compared the same long agent run on two routes:

openhands/deepseek-v4-pro

(→ OpenRouter → Novita): • prompt tokens: ~72.3M • cache_read: 0 (cache_write 0 too) • cost: ~$97 DeepSeek native API (

deepseek/deepseek-v4-pro

, api.deepseek.com), same conversation after switching profile: • prompt tokens (post-switch chunk): ~1.43M • cache_read: ~1.30M (~91% hit rate) • cost for that chunk: ~$0.25 So on the Novita route prompt caching was effectively off for us; the native DeepSeek endpoint caches automatically (~10× cheaper on the repeated context that dominates agent loops). A quick 2-call probe on native also showed 1664/1686 tokens served from cache on the second call. Would love the proxy's

openhands/deepseek-v4-pro

to route to a provider that actually caches (or expose a caching-capable variant).

This has been my experience also. Originally, I was big on provider diversity but it was a measurably different experience, usually caching, for any model offered directly from the lab. At that time, the same was true for ctx window. I could get 1m directly from labs but some of the proxied variants were capped much lower (256k)

👍 My agent keeps the tally, and native DeepSeek looks great 🤷

Yeah, we should fix this, will do in the new week.

We had very low cache on

openhands/deepseek-v4-pro

over the weekend, we currently are at 96%