Tried using Meta's Muse Spark 1.2 model the other day and was pretty surprised checking the bill. Makes you really appreciate the engineering work Deepseek does with caching.

preview.redd.it/7d2cvkqgg8kh1.png Fyi: this was all research agent work which was running pretty much directly after each other. Agents were launched to do tasks a run in a loop for 5 minutes. This flow is predictable for caching usually.