Estimating Costs of AI in Software Development

As AI becomes more entrenched in the development cycle, it’s interesting to think of how much it costs. When we develop with agents, we’re not just paying the upfront token cost for the code itself. We’re paying for the agentic code review scan, the automatic PR summary that gets generated, the planning session needed to create the ticket, and more.

By OpenAI measurements, 1 token is approximately 4 characters. If we assume 1 line of code is 80 characters, we can assume that each line of code costs 20 tokens to generate. With that as the base, we can start to estimate the cost of each layer of AI assistance in the development process.

All things considered, this spend is more modest than I expected. And it seems inline with Claude’s own estimates¹. But let’s look at another view. The more frequently we create pull requests or commits, the more time we’re spending building and deploying software. Consider a deployment pipeline where a new artifact triggers builds of downstream dependencies, deploys to multiple environments, and then runs tests. We can estimate how many compute hours are being spent on builds.

Each build/deployment/test wouldn’t be running on a single tenant machine, but it’s still interesting to think about these second order effects.