Is OpenCode Go a scam? This is what I think might be going on.

Full context on where that's from and a highlevel walkthrough of my session data here

My original post that I mention

Full slop description of the output token difference

I don't think it's a scam in the classical sense but it is a waste of time & money. I fell for it for a few weeks then cancelled

yes opencode models sre nerfed and quantz

Idk honestly and its not my job go do QA for them. I do think doing a shot test or benchmark is missing the actual problems.

Scam...not sure about that, but it has gotten worst across the board, with all models. I'm looking at going full 🐳 harness! Esp if this new super agent design develops. Just seeing if it's worth it.

Superagent looks interesting. DS Bots is also interesting.

You're forgetting that DeepSeek runs on Huawei infrastructure and has its own inference stack, optimised for their hardware. You're not comparing like for like. OpenCode Go should be tested against other third party providers, not DeepSeek's own infrastructure.

I tried 4.1 Flash with Deepseek Harness and the model does perform better there.

Use open code for free

quite speculative for someone with "ai optimist" in their bio

They selling data, thats the only thing thats different. Plus opencode is performing poorly compared to deepseekharness

I use pi-durable in both cases but i hear dsh is better.

@nothingto12 what you think?

It's not a scam. It's just vibe coded slop.

Thank you for this, they're clearly lowing the juice in these models

or .. raising the juice? or its an engineering mistake/bug

oh right. Seems even more strange to me now then

Pi is a good harness working well

It isn't am using it since almost 4-5 months it's the goat

100% with you on this one... Should have also benchmarked CommandCode! I was able to deplete a 5-hour window time in their supposed special plan (GOAT) they got going with DS in less than an hour...

@thdxr

👌 nerfing, qnt models & training is a real thing & this needs legislation! Simple rug pulls is ai 🤖 thing & OC is the KING 👑. GPT GO is a FWOT&£ too, it’s just Luna! The AI 🤖 world needs some reforms soon as Claude is about to IPO to junk bond status too.

@thdxr

M

Ye bc the harness (opencode) is trashing it's context windows (system prompts & bad caching) and the models need to recompensate with more reasoning

It's been proven that DS thinks more on lower quants. So now you know. Compare it with same everything just the provider change and see

I didn't know this! I guess this makes it clear

All I know is I just signed back up to Command Code GOAT, same cost level as Go, much better sleep at night 😆

lol 🤣 yes DS is quantized, i tested it. BUT its good and almost free for bug hunting. for bug fixing GLM is good also(it less quantized).