Evidence correction, 14 October 2026: earlier versions of this post mixed two different Uber adoption metrics, misdated The Information's Uber report, cited Meta's Claudeonomics figures to articles that do not contain them, and misstated what DX, Gartner, Augment Code, GitHub, AI.cc and Pragmatic Engineer reported. They also attributed an arXiv study to Microsoft Research and gave figures no source supports, including a GPT-5 price and a 40% to 60% removable-token share. Those claims are removed or corrected below.
Microsoft is pulling Claude Code. The reason is the bill.
In mid-May 2026, Microsoft began cancelling most internal Claude Code licenses across its Experiences and Devices division. The directive covers thousands of engineers working on Windows, Microsoft 365, Teams, Outlook, and Surface. Access ends June 30, 2026. Developers are being redirected to GitHub Copilot CLI.
The pilot launched in December 2025, and Microsoft encouraged project managers and designers, not only engineers, to experiment with it. The tool was popular. It was also expensive. Microsoft's financial year ends June 30, and the Claude Code line item was hard to defend at enterprise scale.
Microsoft framed the move as platform standardization. The financial pressure tells the rest of the story.
Source: The Verge, "Microsoft starts canceling Claude Code licenses," May 14, 2026
Microsoft is not alone. This is an industry pattern.
Uber burned through its entire 2026 AI budget by April. By March, 84% of Uber developers were agentic coding users, and Claude Code use had nearly doubled from 32% of engineers in December to 63% in February. Per-developer costs ran $500 to $2,000 per month for heavy users. The CTO had to revisit financial assumptions after spending exceeded projections by a wide margin.
Source: The Information, "Uber CTO Shows How Claude Code Can Blow Up AI Budgets," April 2026
Source: The Pragmatic Engineer, "How Uber uses AI for development: inside look," March 2026
Source: Forbes, "Uber Burns Its 2026 AI Budget In Four Months On Claude Code," May 17, 2026
At Amazon, the internal culture went the opposite direction. Employees were encouraged to maximize token consumption in a practice they called "tokenmaxxing." The bet was that heavy AI usage would drive enough productivity to justify the cost. The bill is still being tallied.
At Meta, an employee built an internal leaderboard called "Claudeonomics" that tracked token use across more than 85,000 employees and listed the top 250 users. In a 30-day window, total usage on the dashboard exceeded 60 trillion tokens. Meta shut the leaderboard down after it was reported in April.
Source: Fortune, "Microsoft reports are exposing AI's real cost problem," May 22, 2026
The numbers across the industry tell the same story.
The Pragmatic Engineer newsletter reported one fintech whose token spend rose about 10x in six months, and a seed-stage startup that went from $200 to $3,000 per developer per month across its seven developers.
Source: The Pragmatic Engineer, "The Pulse: token spend breaks budgets – what next?", April 30, 2026
DX surveyed engineering leaders on their 2026 AI budgets. When asked about 2025 spending, 38.4% of leaders reported $101 to $500 per developer per year. For 2026, nearly half planned to allocate 1% to 3% of their engineering budget to AI tools. DX advises budgeting from $500 per developer per year at the low end to $3,000 or more at the high end.
Source: DX, "How are engineering leaders approaching 2026 AI tooling budgets?", October 2025
Gartner's May 2026 report estimates the enterprise AI coding agent market at $9.8 billion to $11 billion annualized.
Source: Gartner, "Leading in Enterprise AI Coding Agents Requires More Than Product Momentum," 2026
Software vendors are also adding their own pressure. AI-driven renewals are raising enterprise software prices 20% to 37% through forced SKU migrations and credit-based pricing. The "AI tax" is compounding on top of the token bill.
Source: Tropic, "Is Your Company Ready for the AI Tax?", December 2025
Why the bill grows faster than the budget.
The root cause is straightforward. Every call in every agentic coding session hits a frontier model. File reads, status checks, linting, formatting, classification, and simple code generation all go to Opus at $5 input and $25 output per million tokens, or GPT-5 at $1.25 and $10. The model does not distinguish between a request that requires deep reasoning and one that Haiku could handle at $1 per million input tokens.
Agentic workflows consume 1,000x more tokens than chat. A study led by University of Michigan and Stanford researchers (arXiv 2604.22750) documented this across eight frontier LLMs on SWE-bench Verified. Runs on the same task can differ by up to 30x in total tokens, and spending more tokens does not reliably improve accuracy.
The math is counterintuitive but consistent. Per-token prices have fallen sharply since 2024. Enterprise AI spend tripled to $37 billion anyway. Cheaper tokens made agentic workflows viable. Volume growth outpaced price cuts.
Source: Menlo Ventures, "2025: The State of Generative AI in the Enterprise," December 2025
Goldman Sachs projects this will accelerate. Their forecast calls for a 24x increase in token consumption by 2030, reaching 120 quadrillion tokens per month. If consumption grows faster than unit costs fall, the bill gets worse, not better.
Source: Goldman Sachs, "AI Agents Forecast to Boost Tech Cash Flow as Usage Soars," May 2026
Source: Bai et al., "How Do AI Agents Spend Your Money?", arXiv:2604.22750, April 2026
The fix is not to use AI less. It is to route each call.
Microsoft's answer was to cancel licenses and consolidate on a cheaper tool. That is one approach. It trades capability for cost control. It does not address the underlying problem: uniform model selection.
The underlying problem is that every API call, regardless of complexity, goes to the most expensive model. File reads go to Opus. Status checks go to Opus. Simple code generation goes to Opus. Many of those calls likely do not need a frontier model, but they get billed at frontier rates anyway.
AI.cc, an API aggregator, reported what happened across its own enterprise accounts. Accounts that fully implemented tiered routing reached a median blended cost of $2.31 per million tokens in 2026, compared with $18.40 for equivalent workloads sent entirely to frontier models twelve months earlier. That is an 87% reduction. AI.cc says the tiered pattern now covers 64% of its enterprise accounts by token volume.
Augment Code published a routing guide for coding agent workflows. Its worked cost model for one hypothetical multi-agent session (Opus for coordination and architecture decisions, Sonnet for implementation, Haiku for file navigation) puts three-tier Claude routing 51% below uniform Opus, $0.98 versus $2.02 per session.
Source: Augment Code, "Best AI Model for Coding Agents in 2026: A Routing Guide," April 2026
What the Microsoft math looks like with routing.
This is an illustration, not a measurement. Microsoft's Experiences and Devices division has thousands of engineers. Assume 3,000 active Claude Code users at an average of $300 per month (conservative, given industry data showing $500 to $2,000 for heavy users). That is 3,000 x $300 = $900,000 per month, or $10.8 million per year.
Now assume 60% of agentic coding calls are low complexity and route to Haiku at $1/$5 per million tokens instead of Opus at $5/$25. If those calls use about as many tokens as the rest, the blended cost drops by 0.6 x 80% = 48%, so call it 40% to 50%. That turns $900,000 per month into $450,000 to $540,000. The annual savings are $4.3 to $5.4 million.
Microsoft chose to cancel licenses instead. The engineers still need AI coding tools. They will use Copilot CLI, which also runs on token-based pricing under the hood. The cost problem did not go away. It moved.
What engineering leaders should do instead.
1. Audit the complexity breakdown. Most teams do not know what percentage of their agentic calls actually need a frontier model. Instrument a week of traffic. Classify each call by complexity. GitHub instrumented its own agentic workflows and cut effective token use by 19% to 62% in the five workflows it optimized.
2. Route per call, not per tool. The model that plans architecture changes is not the model that should read file listings. A trained classifier that evaluates each API call independently and routes to the cheapest capable model captures the pricing gap on the calls that do not need a frontier model. Augment Code's worked cost model puts three-tier routing 51% below uniform Opus for one hypothetical session.
3. Budget with routing in the forecast. DX advises budgeting $500 to $3,000 or more per developer per year, and a heavy agentic user at $2,000 a month passes the top of that range in two months. Budget per developer, and include measured routing savings in the forecast rather than assumed ones.
4. Do not cancel the tool. Fix the model selection. Microsoft's approach trades capability for cost control. Routing preserves the capability (frontier models for tasks that need them) while cutting cost on the calls that do not.
Where Nadir fits.
Nadir can begin beside coding-agent traffic in shadow mode. The Decision API targets about 100 ms and recommends model and reasoning effort together with cache, context, and policy actions. File reads do not automatically belong on a cheap tier; outcome-labelled traces determine which routes are safe to enforce.
Complete non-streaming responses report the served model and actual cost; a configured, priced benchmark adds the avoided-spend comparison. The dashboard aggregates savings by day, week, and month. No instrumentation beyond the two-line change.
For teams watching their AI coding budget evaporate, routing is the most direct lever. It does not require changing tools, rewriting prompts, or downgrading model quality on the requests that actually need it. The requests that do not need a frontier model stop being billed at frontier rates.
Sources: [The Verge, "Microsoft starts canceling Claude Code licenses"](https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad) (May 14, 2026). [Windows Central, "Microsoft cancels Claude Code licenses"](https://www.windowscentral.com/microsoft/microsoft-cancels-claude-code-licenses-shifting-developers-to-github-copilot-cli-a-move-likely-driven-by-financial-motives) (May 15, 2026). [The Information, "Uber CTO Shows How Claude Code Can Blow Up AI Budgets"](https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets) (April 2026). [The Pragmatic Engineer, "How Uber uses AI for development: inside look"](https://newsletter.pragmaticengineer.com/p/how-uber-uses-ai-for-development) (March 2026). [Forbes, "Uber Burns Its 2026 AI Budget In Four Months On Claude Code"](https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code/) (May 17, 2026). [Fortune, "A Meta employee created a dashboard so coworkers can compete to be the company's No. 1 AI token user"](https://fortune.com/2026/04/09/meta-killed-employee-ai-token-dashboard/) (April 9, 2026). [Tom's Hardware, "AI cost crisis hits tech giants as employee tokenmaxxing backfires"](https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-cost-crisis-hits-tech-giants-as-employee-tokenmaxxing-backfires-agentic-ai-eats-up-to-1000x-more-tokens-than-standard-ai-sparks-corporate-pullback-at-microsoft-meta-and-amazon) (May 23, 2026). [Fortune, "Microsoft reports are exposing AI's real cost problem"](https://fortune.com/2026/05/22/microsoft-ai-cost-problem-tokens-agents/) (May 22, 2026). [The Pragmatic Engineer, "The Pulse: token spend breaks budgets – what next?"](https://blog.pragmaticengineer.com/the-pulse-token-spend-breaks-budgets-what-next/) (April 30, 2026). [DX, "How are engineering leaders approaching 2026 AI tooling budgets?"](https://getdx.com/blog/how-are-engineering-leaders-approaching-2026-ai-tooling-budget/) (October 2025). [Gartner, "Leading in Enterprise AI Coding Agents Requires More Than Product Momentum"](https://www.gartner.com/en/articles/enterprise-ai-coding-agent-market) (2026). [Tropic, "Is Your Company Ready for the AI Tax?"](https://www.tropicapp.io/blog/ai-tax) (December 2025). [Menlo Ventures, "2025: The State of Generative AI in the Enterprise"](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/) (December 2025). [Goldman Sachs, "AI Agents Forecast to Boost Tech Cash Flow as Usage Soars"](https://www.goldmansachs.com/insights/articles/ai-agents-forecast-to-boost-tech-cash-flow-as-usage-soars) (May 2026). [Bai et al., "How Do AI Agents Spend Your Money?" arXiv:2604.22750](https://arxiv.org/abs/2604.22750) (April 2026). [AI.cc, "AICC Report: Enterprise Token Costs Drop 67% Year-Over-Year"](https://www.einpresswire.com/article/911544568/aicc-report-enterprise-token-costs-drop-67-year-over-year-as-multi-model-ai-adoption-hits-record-high) (May 2026). [Augment Code, "Best AI Model for Coding Agents in 2026: A Routing Guide"](https://www.augmentcode.com/guides/ai-model-routing-guide) (April 2026). [GitHub, "Improving token efficiency in GitHub Agentic Workflows"](https://github.blog/ai-and-ml/github-copilot/improving-token-efficiency-in-github-agentic-workflows/) (May 2026).