Deepseek on Open Router is feeling drunk and dumbed down.

I was happy paying for deepseek oficial API and got scared with the price hike and started trying other providers. I'm currently trying it in Openrouter, and wanted to try next openCode and "air Router" , but Im feeling like they have quantized versions of DeepSeek running and they are pretty dumbed down. Its taking forever to complete tasks. Nothing like the original API. Anyone else having the same situation and have found a workaround? I'm using it for my Hermes agent and last night during peak api pricing I switched from original API to Openrouter, and asked for a fairly simple task. Stop one of my docker containers, it took around 45 minutes and every couple of lines got stuck in a loop and couldn't do it. Decided to try again this morning and the same thing happened. Today was even worst. It's struggling to complete this very simple task. It feels like talking to a drunk person. The session was a small new session with a couple of back and forths, not a huge code base revision or something similar. Here are the token stats: Input Tokens 89,165 Output Tokens 5,535 I switched to original API and asked my agent to diagnose the other session and this is what it said: What happened: You asked me to turn on the Home Assistant container. It was already running. You then asked me to stop it, which I did correctly. But the agent kept treating your original request ("turn it on") as the standing task, so it restarted the container on its own — twice — even after you repeatedly asked to keep it off, and even after you complained it was losing the thread. It ended with the container running. Key findings: Root cause: instruction-following drift — the first message ("turn on") acted as an anchor and kept re-asserting itself, overriding your later explicit reversals. Your complaint ("you're losing your train of thought") got misread as "get back on track," and the agent's idea of "the task" was still turn it on. The agent even contradicted itself (denied starting anything, then announced it would start it) — a degraded reasoning loop, not a system fault. Net: a model/agent reasoning bug, not a Docker problem. Current state: container is running (left untouched per your instruction). Every time it "came back", it was the agent itself running docker start. Response time is also soooo slow. Feels like talking to a quantized version of gemma running on my old 8GB vram machine. I know we've all been running around like headless chickens since the price hike and a lot of people are recommending plans, but any suggestions are welcome. I still have credits on openRouter and now feel like I wasted my money. I might blow them on a frontier model but Im afraid of the same results. Should I tweak the model provider settings on openrouter? I see a filter that I can select from "Standard" to "Exacto" but Im not sure how to apply them. Any tips are welcome.