The blank-check AI coding era is dead. Here’s what comes next.
The New Stack Amanda Caswell ● Covered by 2 sources
Microsoft just put a leash on its own AI coding tools, giving every division a token budget instead of unlimited access. Turns out letting engineers burn through AI credits doesn't automatically mean better software.
Microsoft spent the last couple years telling engineers to use GitHub Copilot as much as possible. Now it's telling them to slow down and think about the bill. According to an internal memo first surfaced by 404 Media, EVP Jay Parikh has rolled out AI token budgets across the company's divisions, treating coding tokens the way Microsoft treats any other pricey compute resource — something to be tracked, not just consumed. Employees can now see their own token spending, and GPT-5.6 Sol has become the default model inside Copilot.
That default choice is the interesting wrinkle here. Sol isn't the cheap option. OpenAI prices it at $5 per million input tokens and $30 per million output tokens, well above Terra's $2/$12 or Luna's $0.20/$1.20 following July's price cuts. So this isn't a simple story about Microsoft pinching pennies. Parikh's own words in the memo make the framing explicit: "We are not optimizing for fewer tokens. We are optimizing for more impact per token." Translation: spend smart, not less.
There's also a control angle buried in the model-selection story. GitHub Copilot normally routes requests automatically across model families, and CNBC reported that this routing sometimes sent Microsoft's own employees to Anthropic's models. By hardcoding Sol as the default, Microsoft claws back some say over where its internal spending actually lands. That move follows May's decision to strip most Claude Code licenses from its Experiences and Devices division and push engineers toward GitHub Copilot CLI instead — another step toward keeping the whole coding stack, and its data, inside Microsoft's own walls.
The productivity case for all this remains thin, and Microsoft knows it. Its own researchers found that engineers using Claude Code and Copilot CLI merged about 24% more pull requests over four months than expected. That's a real number, but it measures activity, not outcomes — nobody's tracked whether that code shipped fewer bugs or actually mattered to customers. Meanwhile, autonomous agents that explore repositories, run tests, and revise their own work can quietly rack up enormous token bills before a human ever notices something went sideways.
Microsoft isn't alone in learning this lesson the expensive way. Uber reportedly blew through its entire 2026 AI coding budget in four months. Amazon spent $1.8 million on a Claude Sonnet project that missed its budget by 860% and never even shipped. Adobe, Atlassian and Citi are all tightening their own controls now too. The unlimited-token era of AI coding is quietly ending, replaced by something closer to ordinary IT budgeting — with spreadsheets, budget owners, and someone eventually asking what all that spending actually bought.
My take
This is the corporate world rediscovering that unlimited access to a powerful tool doesn't equal good management, a lesson every department eventually relearns with cloud spend, contractor hours, or free snacks in the break room. The real story isn't Microsoft getting cheap, it's Microsoft admitting nobody actually measured whether AI-generated code made anything better, which should worry every company still bragging about token usage as a proxy for progress.
Read more about this at: The New Stack
