Brief experience report on the DeepSeek v4 Flash model

NOTE This is not going to be exhaustive, I didn’t compare them against exactly identical tasks, and this blog post is effectively me turning a somewhat wordy journal entry into a short blog post. Take it with an appropriate amount of salt as per your dietary needs. I’ve recently been testing out the famous Matt Pocock skills package to rebuild a slop tool that had achieved slop singularity and become such a mess not even the clankers could figure it out any more. I’ll probably write a separate thing about that process (early hint: I get why it has 260K stars) but this post is mostly about the swap from Codex to DeepSeek. For access to DeepSeek models, I signed up for OpenRouter and paid them $10 for an initial top-up of credits. I had already built around 70-80% of factual-budget with Codex (last pre-DeepSeek commit is this), and swapped to DeepSeek to implement the final feature of writing out transactions to Actual Budget. I haven’t figured out how to share the session transcript in…