Trying DeepSeek V4 with OpenCode

Regarding my last post: While I spent plenty of thoughts on the whole implementation, most of the coding work was actually done by DeepSeek V4 Flash and Pro, and I just supervised and criticized it.

While I mainly used Amp Free in the last few months, Amp Free is dead. The daily free usage was cut to 5 dollars, and with the expensive prices of Claude Opus, that’s used up pretty fast. GitHub Copilot (Student) isn’t also really useful anymore.

So I discovered OpenCode. There are free models, but it can also be connected with other providers (like Cortects) or with their Go subscription (10 dollars per month; the first month is 5 dollars).

I tried it with Cortects (which is quietly adding many features to become a viable OpenRouter alternative) but also spent 5 dollars on the OpenCode Go subscription to try OpenCode without fear of spending endless money on tokens.

And given DeekSeek V4 (especially Flash) is so cheap compared to the other frontier models, it’s remarkable how capable it is. Most things were coded fine in the first run; I just had to remind it sometimes of all the edge cases to consider and to take special care about things that are important for me (like memory-efficient implementations). Just one time it deleted a file through an edit, restored it with Git, but then deleted it with the same edit again.

With the option to also use these open models with European providers, it’s an even better option, in my opinion. Even though they come from China. OpenAI and Anthropic probably aren’t much better, given the USA are on the edge of fascism.

I think I will not continue my OpenCode Go subscription. Occasionally it happens that OpenCode gets stuck, and I need to interrupt it and type “continue”, while with Cortects it works fine. I just want to see how many tokens I actually need. I guess it might be cheaper to just go with Pay-to-Go using Cortects then.

Regarding AI in general: Looking at my AI-tag archive, it’s mind-blowing to see how rapidly everything seems to develop since 2023. But I also became more critical of mindless AI use and the AI hype again. That’s why I disabled the AI summaries on the blog for now, as they were not providing any real benefits, to be honest. What stays enabled is the aiimage-plugin that can generate image descriptions and alternative texts.