One month coding with GLM 5.3 Flash

Thibaud Colas from the Wagtail core team describes a month-long experiment to restrict development work to the efficient open model GLM 5.3 Flash. The team consumed about 2 billion tokens in September, but only half of that was processed by the target model; the first half of the month stayed within budget (roughly $68, ~4 kWh, ~365 g CO2) while the second half diverted 1 billion tokens to other models. A vibe-coded prototype picked an unsuitable model and burned through 450 million tokens overnight (~$150, ~5 kWh), and platform capacity limits forced temporary switches to peers like DeepSeek V4.1 Flash and Qwen 3.8 Flash. Benchmarking work across 14 models is ongoing; a sneak-peek table ranks DeepSeek V4.1 Flash highest at 95% accuracy, 14.9 Wh energy use and $0.09 per task.

The write-up turns failures into clear, actionable lessons: implement constant, local measurement of tokens, energy and spend; budget explicitly for experimentation as distinct from routine tasks; improve prompt design and multi-agent orchestration with bounded roles; and prioritize more efficient inference techniques and models. The core recommendation is operational: aim for the majority of day-to-day AI inference to run on flash-tier, cost- and energy-efficient models rather than judging efficiency by token counts alone. An invitation to follow progress at Wagtail Space 2026 closes the piece.