One month coding with GLM 5.3 Flash

In September, a challenge was set to dedicate the entire month to only one cost-effective open model, GLM 5.3 Flash. Despite the initial enthusiasm, the experience proved to be more complex than anticipated. A total of 2B tokens were utilized, with 1B of those tokens allocated to other models during the latter half of the month.

Initially, the second half of the month presented challenges, with 1B tokens directed towards models other than GLM 5.3 Flash. However, this early setback was anticipated and accounted for in the planning process.

The project's prototype, a Wagtail MCP server, proved to be a useful demonstration of GLM 5.3 Flash's capabilities. However, it also highlighted the importance of careful model selection and agentic patterns. The experience served as a valuable lesson in budgeting and being mindful of model choices. In future endeavors, the team intends to allocate a larger portion of their AI inference work to efficient models, aiming for a 50% or higher utilization of such models in day-to-day production tasks.

Infrastructure availability issues also posed a challenge, with degradation in performance observed for GLM 5.3 Flash due to its high rank on the Pareto frontier of relevant models. To mitigate this, the team switched to alternative models like DeepSeek V4.1 Flash and Qwen 3.8 Flash, which were readily available in European data centers. This strategic pivot allowed the team to continue their work without significant disruption.

Looking ahead to October, the team aims to refine their approach by focusing on one or two flash-tier efficient models for day-to-day developer work. They believe that most AI inference work should be conducted using such cost-effective models, measured in terms of cost or energy consumption rather than the number of tokens used.

This strategy will help reduce energy consumption and improve overall efficiency. The team is eager to share their learnings and experiences at Wagtail Space 2026 in November, offering insights into how their approach evolved throughout the month-long challenge.