DeepSeek V4 reshapes coding spend

This reads less like a model showdown and more like a task-routing audit. The useful takeaway is that coding work splits cleanly by context size and risk, so the right optimization is choosing the cheapest model that clears the bar for each bucket.

• –File reads, project scans, and "explain this code" were the easiest local wins, which is exactly where latency and cost matter most.
• –Single-file edits and test writing stayed strong locally, and the miss rate sounds reviewable rather than catastrophic.
• –Multi-file debugging is where local starts to sag, which lines up with the need for broader context and more reliable cross-file reasoning.
• –The 5+ file refactor bucket still justifies cloud, but it is a minority of the workload, so a mixed routing policy makes more sense than defaulting everything to frontier models.
• –The post is self-reported and sample-sized, but it makes the economics concrete: the expensive part of the stack should be reserved for genuinely hard work.