Meta Muse Spark 1.3 model is only 3 cents more expensive than qwen-3.7-flash if you agree to let them train on you. Not getting trained on is 14 times more expensive. I really need a super fast model at the moment. This model is extrodinary powerful while having much performance.

Maybe they do even caching right. DeepSeek was primarely so cheap because their cache system was top notch. Recently I did deepseek at open router and the money was just vaporized because their bad caching mechanism compared to deepseek provider itself.