TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A Wagtail team’s month-long effort to use GLM 5.3 Flash for coding succeeded for about half of its token use, but infrastructure constraints, prototyping and research led it to use other models too. The team reported $68 in spending for its first-half use of GLM 5.3 Flash, while total use across the month reached about 35 kWh—above its 10 kWh target.
Wagtail’s month-long coding experiment with GLM 5.3 Flash used the model for about half of the team’s total token consumption, falling short of its goal to make it the sole model for the month. The team said GLM 5.3 Flash worked well for a range of development tasks, but provider capacity problems, prototype costs and research needs pushed usage toward other models and raised total energy consumption to about 35 kWh, against a target of 10 kWh.
In a retrospective published on the Wagtail CMS blog, the team said it spent the first half of the month using GLM 5.3 Flash, with that period’s use costing $68 and accounting for about 4 kWh of energy and 365 grams of carbon emissions, according to its estimates. During the second half, the team used other models, which received about 1 billion tokens. Across the month, it said the target model handled 1 billion of roughly 2 billion tokens.
The team also described a separate cost tied to its experimental Wagtail MCP server, a prototype built using a “vibe coding” approach. It said choosing a model that was not well suited to the work led to about 450 million tokens being used almost overnight, costing $150 and consuming an estimated 5 kWh. Wagtail said the server worked well as a demo, while noting that similar results might have been achievable at a substantially lower cost with more effort.
Provider availability became another constraint. Wagtail reported degraded performance for GLM 5.3 Flash and said it switched to DeepSeek V4.1 Flash and Qwen 3.8 Flash. It said the comparison was limited to open models it could confirm were available in a European data center; the available alternatives and trade-offs could differ for teams using providers or regions outside that scope.
The Cost of Coding With AI
The retrospective offers a practical account of the gap between choosing an efficient model and keeping a real development workflow on it. Wagtail’s experience suggests that a model’s price and capabilities are only part of the equation: provider capacity, task design and experimentation can all affect actual usage and cost.
That matters for organizations trying to manage the financial and environmental costs of AI-assisted development. Wagtail reported both energy estimates and spending, rather than treating token counts alone as a measure of efficiency. Its total of about 35 kWh versus a 10 kWh target shows how prototypes and model switching can change the outcome, though the report does not establish that the same figures would apply to other teams or workloads.
The team said GLM 5.3 Flash remained suitable for day-to-day work, and argued that developers could direct a majority of routine inference toward cheaper “flash-tier” models. That is a recommendation based on Wagtail’s own experience, not an independent comparison proving those models are best for every coding task.
As an affiliate, we earn on qualifying purchases.
How the Month Was Measured
Wagtail described the effort as a month-long attempt to focus coding work on GLM 5.3 Flash. The blog report says the model was used for work on Wagtail itself, websites built with the CMS, interface tasks, AI research and development, documentation and evaluations. The team highlighted the model’s 1 million-token context window, vision support and availability through multiple providers as useful features.
The experiment also included activity beyond routine engineering. Wagtail said it needed to test a wider range of models for its work benchmarking performance on Wagtail tasks. It is developing a benchmark intended to provide concrete comparisons that could help guide model selection, alongside agent skills and a command-line prototype designed to work with agents.
The team’s report distinguishes its intended production use from research and development: it wants to focus future efforts on using an efficient model for more than half of normal day-to-day work, while budgeting separately for experimentation. Its estimates and conclusions are based on its own workload, provider access and accounting method.
““GLM 5.3 Flash itself was excellent!””
— Wagtail team
As an affiliate, we earn on qualifying purchases.
Limits of the Usage Figures
The report does not provide a full breakdown of energy use, spending and task outcomes for every model used during the month. It also does not explain in detail how energy and carbon estimates were calculated, so readers cannot directly compare those figures with estimates made under other methods.
Wagtail attributed GLM 5.3 Flash’s performance degradation to likely capacity constraints, but did not identify a single confirmed cause or quantify how much availability affected the results. Its model comparison was restricted to options it could confirm were available in European data centers, meaning the findings do not describe the full market.
The team’s suggestion that the prototype could have reached similar results for about five times less cost is an estimate, not a reported controlled comparison. The blog does not specify which alternative model or workflow would have produced that result, or how the team assessed equivalence.
As an affiliate, we earn on qualifying purchases.
Wagtail’s October Efficiency Target
Wagtail said it plans to repeat the effort in October, this time focusing on using efficient models for more than half of its normal day-to-day production work rather than trying to exclude other models altogether. It said it intends to track usage, spending and energy more consistently and to set aside a budget for experimentation.
The team also plans to make more deliberate decisions about which prototypes to build and to test different agent roles, including orchestrator, scout, implementer and reviewer. It said it wants to continue evaluating more efficient models and use benchmark data to guide choices. Wagtail said it expects to discuss the effort at Wagtail Space 2026 in November.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did Wagtail use GLM 5.3 Flash for the entire month?
No. The team said it used GLM 5.3 Flash for the first half of the month. Other models received about 1 billion tokens during the second half and other work, leaving the target model with about half of total token use.
How much did the GLM 5.3 Flash use cost?
Wagtail reported that the model’s use during the first half of the month cost $68 and used about 4 kWh of energy, with estimated carbon emissions of 365 grams. That figure is not the total cost of the month’s AI work.
Why did the team switch to other models?
Wagtail reported performance degradation and availability constraints for GLM 5.3 Flash. It said it switched to DeepSeek V4.1 Flash and Qwen 3.8 Flash, while also using other models for research and development.
What was the MCP server prototype cost?
Wagtail said the prototype used about 450 million tokens almost overnight, costing $150 and consuming an estimated 5 kWh. The team said the server worked well as a demo but that its model choice and workflow could have been more efficient.
What is Wagtail changing for its next experiment?
For October, the team said it will focus on using efficient models for most day-to-day work, while treating research and prototyping as separately budgeted activities. It also plans more consistent tracking of cost, energy use and outcomes.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
