As a reminder, both the web interface (chat.deepseek.com) and the API (for agentic work on your machine) were upgraded to the full deepseek 3.2 recently. This was done on deepseek’s part and there is no way to go back to the previous model.

I’m using it to write some scripts right now, and it definitely works. Even without the LSP being activated it does very little wrong work. You just throw it at the problem and let it do its thing, even if it may seem nonsensical at first. Even for more ‘complex’ problems e.g. giving it like 3 different things to fix that have no similarities, it will eventually get to it.

I also notice that once you’ve made a local git repo for your project it will happily commit to it at the end. Haven’t had to work on the repo myself (e.g. rolling back to a previous commit) so I can’t say much about it, but it’s nice that it does it by itself even on subsequent context windows.

However as always the problem remains context window and token wastage. The second one means it’s more expensive to use deepseek for agentic workflow - I’m already through 5$ worth of tokens, though granted it’s not a super simple script (lots of external tool calls). The prices haven’t changed, deepseek API is still as cheap as ever, it’s just that it needs a lot more tokens to work.

Context window is alright. This is generally a bottleneck but crush handles it gracefully - when the window reaches 80% (could be 90% tbh, not sure if I can edit that) the instructions are cut and the LLM is instructed to summarize the context for this session. Then crush automatically resends your last prompt with the added “the last session was cut off because of context window size, the user’s prompt was: [prompt]” and the LLM picks back where it left off.

Except that it starts over from your last prompt, including checking what it did in the last session, and 30% is already pre-filled (I think because both crush and the API send a system prompt), so you just keep hitting the limit over and over again, and this also sends a lot of tokens and therefore costs money.

It’s confusing at first because you think the LLM is just spinning in circles, but actually if you throw it at the problem and let it crunch the numbers it will eventually get the result. I only hope they can do something about the token usage because this is more expensive than the previous version still, if you calculate per task.