A month trying to code with only GLM 5.3 Flash

Thibaud Colas, who works on the Wagtail CMS, gave himself a simple rule for September: do all of his AI-assisted coding with one cheap open model, GLM 5.3 Flash, and nothing else. He did not quite manage it. Over the month he burned through about 2 billion tokens, but only around half of those went to the model he meant to test. The rest leaked to other models whenever a task felt too hard or an API misbehaved, which is itself the honest finding: single-model discipline is difficult once real work gets in the way.

What lifts the post above a diary entry is that he measured everything. The focused first half of the month cost about $68, and he tracked energy as well as money, where the surprise was unpleasant: actual energy use came in around 35 kWh against a rough projection closer to 10 kWh. For comparison he cites DeepSeek V4.1 Flash hitting 95% accuracy on his task set at roughly $0.09 and 14.9 Wh per task, a reminder that a model's headline price is not the whole bill. His write-up ends with practical fixes for a second attempt: better routing between models, a real token budget, and sharper prompts to cut wasted calls.

Why it matters

If your team is trying to standardize on one coding model to control spending, this is a useful reality check. Sticking to a single model is harder than it sounds, and energy, not just the dollar figure on the invoice, is worth measuring before you commit.