DeepSeek V4.1 Flash 1017: direct API beats OpenCode free tier on cache and cost
A user finds direct DeepSeek API gives 99.6%+ cache hits versus ~98% on OpenCode free, which compresses context from ~300K and runs slower; OpenCode Go turns $10 into ~$80 of inference.
DeepSeek V4.1 Flash 1017 taught me something i didn’t expect: the same model can feel completely different depending on where you run it. 1/ direct API feels noticeably better than OpenCode free for me. the free route is slower and seems to make the model do more unnecessary work. 2/ cache quality matters more than i expected. in OpenCode i’m seeing around 98% cache hit, while direct DeepSeek can regularly push 99.6%+, and on long sessions even close to 99.95%. 3/ the free context workaround is basically dead. i used to be able to stretch sessions close to 1M tokens, but now compression starts around 300K. 4/ on OpenCode Go, the economics are still hard to beat: $10 gives roughly $80 worth of inference, so i can run DeepSeek for much longer sessions without thinking about every token. 5/ 1017 also feels more efficient than the previous checkpoint. the price is the same, but it loops less, so i’m spending fewer tokens to get to the same finished task. provider, cache and context setup matter way more than i expected.
This post quotes another post: “DeepSeek V4.1 Flash 1017 in OpenCode: fewer loops, usable vision”