Just try. In a case last week it was ~3x and I tried multiple providers: deepseek, gmicloud/fp8, novita/fp8, and another one I can't remember. It was a large job where at least 2/3rds of the start of the prompts was exactly the same (literally a static string).
Then I read somewhere (I think X) that OpenRouter adds stuff and breaks caching (telemetry? headers? can't remember). So I stopped the job, switched to actual DeepSeek provider, and voilá, caching 3x more tokens per request (on average).
Comments
Just try. In a case last week it was ~3x and I tried multiple providers: deepseek, gmicloud/fp8, novita/fp8, and another one I can't remember. It was a large job where at least 2/3rds of the start of the prompts was exactly the same (literally a static string).
Then I read somewhere (I think X) that OpenRouter adds stuff and breaks caching (telemetry? headers? can't remember). So I stopped the job, switched to actual DeepSeek provider, and voilá, caching 3x more tokens per request (on average).
I meant actual DeepSeek API.