The lede
Why isn't the industry freaking out about DeepSeek 4.1 Flash?
dgt.is
A month of daily use across a dozen projects, and the claim is not that the model is better — it is that mid-session the author cannot reliably tell it from Opus, at a price that breaks the mental model everyone is budgeting against. The example to argue with is a task that runs about $1 on a frontier model coming in near $0.003, with all-day sessions staying under a dollar on a $10 OpenCode subscription. The mechanism he credits is the KV cache: DeepSeek reportedly shrank it by something like 437x since V1, and since that cache lives in GPU memory and grows with every hour of a long coding session, shrinking it is exactly where the money in agentic work goes. His workflow is the practical part — the cheap model does the planning, the exploration and the final fixes, and Opus 5.5 gets used for a second opinion rather than for raw capability. Worth reading against your own bill before the next renewal.