Cheap is not the same as free
The lowest invoice is a bad model that retries ten times. AICoder prices are vendor list rates per million tokens. DeepSeek-V4-Flash at $0.22 input / $0.66 output is the floor among models we still treat as coding-capable. V4-Pro is the cheap model you can leave in an agent loop.
Qwen and Luna as alternatives
Qwen3.8-27B ($0.30 / $0.90) is the open-weight-adjacent option if you want a Chinese-lab stack. GPT-5.6 Luna is OpenAI’s cheaper tier when you must stay in that API. Neither replaces V4-Pro for raw dollars-per-merged-PR if your tasks are long-horizon agents.
Where cheap models break
They drop steps in multi-file refactors, invent APIs, and need tighter tests. Pair them with Cline, Aider, or Continue so the loop can compile and retry. If retry tokens exceed the Opus premium, you picked the wrong tier.