Stop counting tokens, start counting what ships
Original reporting by MIT Technology Review

AI coding agents refer to AI systems that autonomously plan, generate, test, and correct code, far exceeding basic chat interactions. While incredibly powerful, their iterative nature—where a single task can consume 10 to 50 times more tokens than basic prompts—poses a growing economic challenge. Heavy usage is already driving companies like Uber to exhaust budgets and GitHub Copilot to shift pricing models. Analysts forecast AI token costs may soon exceed developer salaries.
The initial industry response of rationing usage caps, however, is proving counterproductive. These blanket limits optimize for consumption, not results, often stifling valuable workflows and failing to differentiate between waste and productive work.
Rethinking Cost
A more effective approach measures true output: shipped software. By tracking "cost per merged pull request"—total AI-assisted development cost divided by changes integrated—organizations connect AI spending directly to tangible results. This demands a strategic shift: reserving premium models for complex reasoning, while routing routine tasks like code generation to more economical, often open-weight, self-operated alternatives. Leveraging owned infrastructure for high-volume tasks further reshapes economics, turning incremental costs into efficient, upfront investments. This hybrid approach, demonstrated by significant token cost reductions, enables widespread AI adoption without sacrificing budget control or data sovereignty. The goal: an operating model where increased AI use improves, rather than undermines, economic efficiency.
The rapid ascent of AI agents in software development, while promising unprecedented productivity gains, has simultaneously unveiled a critical challenge: the escalating token costs threatening to undermine their economic viability. Simply rationing usage is a shortsighted response, as it fails to differentiate between wasteful consumption and genuinely value-generating activity. Instead, the industry must pivot its focus to measurable outcomes, prioritizing metrics like cost per merged pull request to ensure AI investments directly translate into tangible, high-quality software. This shift necessitates a sophisticated architectural strategy: intelligently matching tasks to models, reserving premium frontier AI for complex, novel reasoning, and routing routine code generation, testing, and refactoring to more cost-effective, often self-operated, open-weight models.
Shaping Future AI Development
Crucially, the economics of AI-assisted development are poised for a significant transformation through the strategic adoption of owned inference capacity. By internalizing high-volume, repeatable workloads, enterprises can convert unpredictable variable expenses into more stable, scalable operational costs, while simultaneously gaining invaluable sovereignty over sensitive intellectual property. This emerging hybrid model, seamlessly blending external frontier AI with internal, owned infrastructure orchestrated by intelligent gateways, will define the next era of enterprise AI. Organizations that master this nuanced approach will not only control their spend but also unlock profound competitive advantages in development velocity and innovation. This re-architecting transforms AI from a potential financial drain into a powerful, secure, and economically sustainable force multiplier, fundamentally reshaping how software is built and accelerating digital transformation across industries.
Frequently asked questions
- Why are AI coding agent costs becoming a significant challenge for enterprises?
- AI coding agents automate complex tasks like planning, code generation, and testing, often repeating steps. This iterative process consumes 10 to 50 times more tokens than basic chat prompts. With per-token pricing, high usage for valuable workflows rapidly escalates costs, straining budgets and necessitating new strategies for economic efficiency in software development to avoid overspending.
- What is a better way for companies to measure the value of AI in software development?
- Instead of tracking consumption metrics like cost per token, a more effective measure is "cost per merged pull request." This metric divides total AI-assisted development cost by the number of code changes successfully reviewed and integrated into the main codebase. It focuses on finished work, allowing organizations to ensure AI investment directly translates into shippable software, considering quality and rework alongside cost.
- How can organizations reduce the high costs of using AI coding agents?
- Organizations can reduce AI coding costs by adopting a hybrid strategy. This involves matching the AI model to the task, using expensive frontier models only for complex reasoning and routing routine work to more cost-effective open-weight or on-premises models. Additionally, owning computing capacity for high-volume tasks makes each call cheaper over time, offering both cost savings and enhanced control over sensitive code.