AI
Major tech firms tighten spending controls as AI coding tools drain budgets
After years of unfettered AI adoption, companies like Microsoft, Amazon and Uber are now tracking token consumption and enforcing budgets to ensure their coding agents deliver real value.
The New Stack
Enthusiasm for AI-powered coding tools has given way to hard financial reality. Tech giants are discovering that simply deploying these systems doesn't guarantee productivity gains — and the token costs can spiral quickly.
Token budgets arrive at Microsoft#
Microsoft has become the latest company to impose spending guardrails on its AI tool usage. The software giant introduced AI token budgets for each division and equipped employees with dashboards to monitor their own consumption. Engineers are directed to default to OpenAI's GPT-5.6 Sol model in GitHub Copilot, though other options remain available.
"Tokenmaxxing is not what we are optimizing for," Microsoft executive vice president Jay Parikh stated in an internal memo. "I want all of us focused on maximizing outcomes that move the needle for our customers and our business." The budgets took effect in July, though Microsoft has not disclosed their size or stated whether access will be automatically revoked at budget limits.
The paradox of model selection#
The shift to GPT-5.6 Sol as the default carries a strategic twist. Despite being positioned as a cost control measure, Sol remains pricey at $5 per million input tokens and $30 per million output tokens — higher than OpenAI's cheaper Terra and Luna alternatives. This suggests Microsoft's real goal is not raw cost reduction but extracting greater value from each token spent.
By consolidating around a single default, Microsoft regains visibility and control over where spending flows. The company simultaneously began phasing out Claude Code licenses in some divisions, pushing engineers toward GitHub Copilot instead. This migration keeps token consumption within Microsoft's own ecosystem and policy framework.
The hidden costs of agentic AI#
As coding tools evolve beyond simple autocomplete, the spending problem intensifies. Agentic systems that explore repositories, draft implementation plans, run tests and revise their work accumulate tokens across multiple steps. An agent that rereads a large codebase or pursues a failed approach can consume significant resources before a developer intervenes. GitHub acknowledged this pressure when it shifted Copilot to usage-based billing in June and introduced user-level budget controls for organizations.
