WSJ is wrong. Autocomplete code models are small (5B param) and very cheap. A single inference is roughly $0.000001.
Continue[0] for example can use up to 3 different models in a session: a large model such as GPT-4 Turbo for chat and code QA, a smaller low latency model such as StarCoder2-3B for code completion, and yet another model such as all-MiniLM-L6-v2 for generating embeddings.