is OpenCode Go a scam?
I pulled every session out of pi-durable and compared 31 OpenCode Go sessions against 108 DeepSeek API sessions.
Other people have reported the same "it feels dumber" thing, and the benchmarks people have run show OpenCode Go and DeepSeek are practically indistinguishable. I believe those benchmarks and I still think something is going on: cold starts and cache hit rates, and caching is not a boolean. The real finding was 51.2% against 39.2% on reasoning. Error rates actually looked better on OpenCode Go.
I also take back the word "quantized." The OpenCode team insists they don't quantize the models, and I have no reason to disbelieve them. I still use OpenCode Zen and will keep using it, but on bang for buck I recommend the DeepSeek API, at least compared to the $40 OpenCode Go tier I was on.