AI Coding Is Getting Faster and More Expensive At The Same Time
AI coding tools started as autocomplete. Developers typed a function and software suggested the next lines. The leading tools now read repositories, modify multiple files, run tests, inspect failures and keep working until they reach a result. JetBrains reported that 90 percent of professional developers in its 2026 survey used some form of AI coding agent at work at least weekly, with 68 percent using them daily.
Usage inside AI companies shows how far the workflow has moved.OpenAI disclosed internal figures indicating that some of its heaviest researchers spend more than $7,000 a day on coding-agent tokens. Median daily spending among researchers rose from $162 in July to more than $600 by mid-August, according to reporting on the data.
The agent is doing more work because the user is asking it to do more work. A developer who once requested a code suggestion now gives the system a problem and lets it inspect the codebase. The agent searches files, writes changes, executes tests and starts another pass when something fails.
Every pass consumes inference. The comparison between coding tools therefore no longer ends with the monthly subscription price.
Two products charging similar seat fees might generate very different model usage once a team turns on autonomous workflows. Some plans bundle usage. Others meter premium models or impose limits. Teams connecting their own model APIs see the cost directly.
Developers face another variable: review. An agent producing 2,000 lines of code in twenty minutes saves little if an experienced engineer spends the afternoon discovering why the change broke a dependency elsewhere. Smaller well-defined tasks usually give both the model and reviewer clearer boundaries.
The highest-value workflow often moves the developer away from typing without removing engineering work. Someone still defines the behaviour, checks architecture, reviews the diff and decides whether the change belongs in the product. Faster code production moves the bottleneck towards judgement.
The tools themselves reflect that change. Claude Code, Cursor, Codex, GitHub Copilot and a long list of newer products now compete over repository context, terminal access, parallel agents, testing, code review and integration with development infrastructure.
Autocomplete has become the least interesting feature. Model choice has moved into the workflow as well. Teams route simple work through cheaper models and reserve stronger models for harder debugging or planning. Developers switch between products according to the codebase and task rather than treating one assistant as the universal answer.
Local execution has gained attention for another reason. Companies do not always want proprietary code travelling through external model services, and predictable high-volume workloads alter the economics of running models closer to the developer.
The market therefore resembles cloud computing more each month. Developers care about capability, but engineering managers also care about usage, security, latency, observability and cost.
AI coding has already crossed the threshold where measuring adoption tells little. A team with 100 developers using an agent every day has not established whether it ships software faster, fixes more bugs or produces more code that somebody later removes. Count completed engineering work instead.
