One developer’s month-long experiment with 2 billion AI tokens reveals why the future of affordable AI depends on smarter model selection, not simply more powerful technology.

WHAT’S HAPPENING

A month-long experiment by a developer on the Wagtail CMS core team has exposed an important reality about artificial intelligence: inexpensive AI models can perform serious development work, but poor model selection and uncontrolled experimentation can quickly erase the savings.

In September 2026, Wagtail developer Thibaud Colas attempted to complete an entire month of software development using GLM 5.3 Flash, an efficient, open-weight AI model.

The objective was straightforward: determine whether an inexpensive AI model could handle everyday software development without relying on more costly alternatives.

The results were revealing.

During the month, the developer processed approximately 2 billion AI tokens, the units AI systems use to process and generate information.

Half of those tokens were processed using GLM 5.3 Flash. The other half went to alternative models after technical difficulties and experimental projects disrupted the original plan.

The first two weeks demonstrated the potential. GLM 5.3 Flash handled development tasks for approximately $68 in reported model-usage costs, consuming roughly 4 kilowatt-hours of energy.

But an experimental AI-assisted coding project later consumed 450 million tokens and approximately $150 almost overnight.

The developer estimated that much of the same work could likely have been accomplished for about one-fifth of that expense through better model selection.

The experiment was documented in an October 2, 2026, report published by Wagtail.

WHY IT MATTERS

The lesson extends far beyond computer programming.

Businesses are increasingly using AI to answer customer questions, develop software, automate administrative work, analyze documents, and operate digital assistants.

Every AI interaction consumes computing resources, and different models carry different costs.

Using the most powerful available AI model for every assignment may be unnecessary, particularly when less expensive models can produce acceptable results.

Wagtail’s experiment demonstrated that efficient models can handle substantial portions of everyday development work.

GLM 5.3 Flash offered several capabilities that made it useful, including a one-million-token context window, visual processing capabilities, and availability through multiple infrastructure providers.

The model was successfully used for website development, interface work, documentation, and AI research.

However, infrastructure congestion sometimes reduced performance, forcing the developer to switch to alternatives such as DeepSeek V4.1 Flash and Qwen 3.8 Flash.

That highlights another business consideration: inexpensive technology only delivers value when it remains dependable.

For organizations investing in AI, the financial opportunity is not necessarily choosing the cheapest model. It is assigning the appropriate model to each task while monitoring costs, performance, and reliability.

The experiment also illustrates why businesses need to distinguish ordinary operating expenses from research and development costs.

A relatively inexpensive system can generate substantial expenses when experimental AI agents repeatedly process information without carefully defined limits.

WHO BENEFITS

Small Businesses: Efficient AI models can make automation, customer assistance, document processing, and software development more affordable.

Software Developers: Lower-cost models can handle routine programming assignments, allowing developers to reserve more expensive systems for demanding problems.

AI Service Providers: Companies offering managed AI assistants can improve operating efficiency by matching different AI models to different customer requirements.

Open-Model Providers: Demonstrations of practical performance strengthen the market position of alternatives to expensive proprietary AI systems.

WHO LOSES

Businesses Using One Model for Everything: Organizations that rely exclusively on premium AI systems may incur unnecessary costs for routine tasks.

Unmanaged AI Operations: Companies without usage monitoring or spending controls risk unexpected expenses from automated processes.

Premium-Only AI Strategies: Providers competing primarily on access to expensive models could face increasing pricing pressure as efficient alternatives improve.

Organizations Without Backup Options: Businesses depending on one AI provider may experience disruptions when infrastructure capacity becomes constrained.

WHAT HAPPENS NEXT

Wagtail’s findings point toward a more disciplined approach to using AI.

The developer identified several priorities for future work: continuous monitoring of AI spending and energy consumption, separate budgets for experimentation, improved coordination among AI agents, and greater reliance on efficient models for everyday tasks.

The broader industry implication is significant.

As AI systems become more deeply integrated into business operations, companies will increasingly need to manage AI resources as carefully as they manage staffing, software subscriptions, and other operating expenses.

This could encourage greater adoption of layered AI systems, where less expensive models handle routine activities while more capable models are reserved for specialized assignments.

There is also an important measurement shift underway.

Counting how many tokens an AI system processes says little about whether the work creates meaningful value.

Businesses ultimately need to measure the cost of successfully completing a task, not simply the volume of AI activity.

Wagtail’s experiment did not achieve its original goal of using one model exclusively throughout the month. Total estimated energy consumption reached approximately 35 kilowatt-hours against a target of 10.

Nevertheless, the results demonstrated that affordable models can perform valuable production work.

The bottom line: The next competitive advantage in AI may not belong to the business using the most powerful model. It may belong to the business that knows when a less expensive model is good enough.

That distinction could determine which companies turn AI into a profitable operating tool—and which simply accumulate another technology expense.