The blank-check AI coding era is dead. Here’s what comes next.

For months Microsoft's line to its own engineers was simple: use Copilot, use it a lot, and don't worry too much about the meter. That era just ended. Every division inside the company has had an internal AI token budget since July, and employees can now watch their own consumption in real time. Jay Parikh, an executive vice president, laid out the shift in a memo obtained by 404 Media, telling staff plainly that "tokenmaxxing is not what we are optimizing for." The goal now, he wrote, is maximizing outcomes that move the needle for customers and the business, not racking up usage for its own sake.

What makes this more than a simple cost-cutting memo is the model Microsoft picked as the new default inside GitHub Copilot: GPT-5.6 Sol, OpenAI's priciest option in that family at $5 per million input tokens and $30 per million output tokens. Compare that with Terra, which OpenAI dropped to $2 and $12 after price cuts on July 30, or Luna at a mere $0.20 and $1.20. So this isn't purely about spending less. Parikh's own words frame it that way: "We are not optimizing for fewer tokens. We are optimizing for more impact per token." Teams now have to weigh capability against price for each specific job rather than defaulting to whatever tops a benchmark chart.

GitHub, meanwhile, has been pulling in a slightly different direction with automatic model selection, letting Copilot route tasks across model families based on the job, subscription tier, and admin policy — a system GitHub says avoids unnecessary costs by working around cache boundaries. But that same automation had, according to CNBC, sometimes routed Microsoft's own employees toward Anthropic's models. Locking in Sol as the default hands Microsoft more control over exactly where its internal token dollars land. It fits a pattern already underway: in May, Microsoft began pulling most Claude Code licenses out of its Experiences and Devices division and pushed engineers toward GitHub Copilot CLI by the end of June, even though Claude remains reachable through Copilot itself.

Microsoft does have some data suggesting the spending isn't wasted. A study from its own researchers tracked tens of thousands of engineers adopting Claude Code and GitHub Copilot CLI, and found they merged roughly 24% more pull requests than expected over the four-month window. That's a real number, but it's also a narrow one — the study didn't check whether that extra code shipped with fewer bugs, tighter security, or any actual value to customers. Merging more pull requests is not the same as building a better product, and agentic tools that read entire codebases or chase a dead-end approach before a human steps in can quietly burn through tokens well before anyone notices.

Microsoft isn't alone in learning this lesson the expensive way. Uber reportedly burned through its entire 2026 AI coding budget in the year's first four months and has since leaned on cheaper default models, caching, and better spend visibility. Amazon's cautionary tale is sharper still: an internal project meant to match author records to product listings using Claude Sonnet reportedly cost $1.8 million, 860% over its planned budget, went unnoticed for months, and never shipped. Two more internal projects reportedly blew past their budgets by a combined $675,000. Adobe, Atlassian, and Citi are now rolling out their own guardrails too, which suggests the industry has quietly moved past the phase where using more AI counted as progress on its own.