The tools that promised to make developers ten times more productive are now generating bills that no one budgeted for. A reckoning is quietly arriving inside some of the world's biggest tech companies.
There is a story that was supposed to go differently. You deploy AI coding assistants across your engineering teams. Productivity soars. Developers ship twice as fast. The CFO is pleased. The story ends well.
What is actually happening at a growing number of companies looks quite different. The tools are being used enthusiastically, sometimes breathtakingly so, and the bills are arriving faster than anyone anticipated. A financial reckoning around AI coding costs is now quietly underway inside some of the most AI-committed organisations on the planet.
In December 2025, Microsoft gave thousands of its employees access to Claude Code, Anthropic's command-line AI coding agent. Engineers, product managers, designers, teams across Windows, Microsoft 365, Outlook and Surface all got in. The idea was to accelerate workflows and reshape how people built software. It worked. Perhaps too well.
By May 2026, Microsoft was cancelling most of those internal licences. According to reporting by TechRadar, the company has told affected teams to shift to GitHub Copilot CLI instead, with the transition expected to complete by June 30, 2026, the close of Microsoft's fiscal year.
The problem was not low adoption. It was the opposite. Unlike traditional software licences where you pay a fixed seat fee regardless of how much an employee uses the product, most frontier AI coding tools are priced on token consumption. Every prompt, every code review, every debugging session, every generated function generates tokens. And tokens cost money. When thousands of engineers use these systems continuously across an eight-hour working day, the numbers compound fast.
"The first wave of enterprise AI adoption was driven by enthusiasm about what the tools could do. The second wave, now underway, is being shaped by what they cost." — TechRadar, May 2026
Microsoft is not alone in this. The story at Uber is even starker. Praveen Neppalli Naga, Uber's Chief Technology Officer, told The Information in April that the company had already exhausted its entire planned 2026 AI coding budget. In four months. The company has roughly 5,000 engineers and by March, Claude Code usage had climbed from 32 percent to 84 percent of that organisation.
What individual engineers at Uber were spending per month on AI coding tokens alone — before the budget ran dry.
Around 70 percent of code committed at Uber now originates with AI assistance, and roughly one in ten live backend updates is shipped by an agent with no human in the loop. The scale of integration made the cost spiral almost inevitable — once a tool is that embedded in daily work, usage does not plateau.
What makes this difficult to manage is not simply that AI tools are expensive. It is that their pricing architecture was designed for a world of occasional use, not the always-on, deeply embedded workflows they have become. Traditional SaaS was linear. Pay for seats, predict spend, done. Token-based billing is nonlinear and exponential.
Standard chatbot interactions consume a relatively predictable number of tokens. But agentic AI systems, which are increasingly what enterprise coding tools are becoming, work very differently. An agent monitoring a codebase, writing tests, running verification loops and handling pull requests autonomously can consume five to thirty times more tokens per task than a simple code suggestion. Each API call carries the full conversation history, so by turn 50 of an agentic session the system is resending every prior message, tool call and result with every new request. The token count grows geometrically as sessions deepen.
Enterprise AI deployment audits have consistently found that retry logic, context window management and retrieval augmentation add between 40 and 60 percent on top of whatever token costs finance teams are already tracking. Most finance teams have no mechanism to track this in real time.
Beyond raw token consumption, a 500-developer team using GitHub Copilot Business already faces roughly $114,000 in annual licensing costs. The same team on Cursor's business tier would pay around $192,000. Tabnine Enterprise exceeds $234,000. But those headline figures are just the floor. Implementation costs for monitoring, governance and internal tooling run anywhere from $50,000 to $250,000 annually on top.
A reasonable response to all of this is to wait it out. Token prices are falling. Inference is getting cheaper. Surely the economics will sort themselves out as the underlying technology matures. Gartner projects that by 2030, inference costs for trillion-parameter models will fall by nearly 90 percent compared to 2025 levels. On its face, that sounds like the problem resolves itself.
It does not. Gartner has been explicit that cheaper individual tokens will not reduce enterprise AI bills, because the volume of tokens being consumed is growing far faster than the price per token is falling. Will Sommer, a Senior Director Analyst at Gartner, has stated that enterprises should not confuse the deflation of commodity tokens with actual democratisation of cutting-edge inference capabilities.
Goldman Sachs forecasts that agentic AI will drive a 24-fold increase in global token consumption by 2030, reaching roughly 120 quadrillion tokens per month. By 2040, that figure could rise to 55 times current levels. Enterprise agents are expected to account for over 70 percent of all token usage by 2040.
The deeper issue is structural. Token-based billing, which is now the default model for every frontier AI coding API, was designed to be intuitive for small experiments and developer prototypes. Pay only for what you use. It sounds fair. But it creates a catastrophically poor fit for enterprise budget cycles.
Enterprise procurement was built around predictable, linear costs. Seat licences. Annual contracts. Fixed infrastructure. The finance team knows what 500 seats of Salesforce costs in January and in December. It knows nothing of the kind with token-based AI. One engineer who shifts from writing code manually to relying on an agentic coding tool could increase their token consumption by a factor of 1,000 in a single month.
Compounding the problem is the extraordinary variance in model pricing. The cheapest production-grade language models in 2026 cost around $0.04 per million tokens. The most expensive frontier reasoning models cost upward of $180 per million tokens. That is a 4,500-fold spread between the cheapest and most capable options. Most enterprise deployments have no automated logic to route simpler tasks to cheaper models.
The market response is still forming, but a few patterns are becoming visible. First-party consolidation is one route. Microsoft's move to push teams toward GitHub Copilot CLI is partly about cost, but also about platform strategy. When a company can control the tool and negotiate pricing as the underlying model provider, the budget dynamic changes.
Real-time cost visibility is emerging as a new procurement requirement. The question being asked in procurement conversations now is not just whether an AI coding tool is productive but whether the organisation can monitor, forecast and cap token spend in real time before deploying at scale.
Model routing is the other significant lever. Intelligent infrastructure that automatically routes low-complexity requests to cheaper models and reserves premium inference for tasks that genuinely require it can reduce enterprise AI spending by 40 to 70 percent. The average organisation's OpenAI API spend reached $384,500 annually as of April 2026, and most of that spend carries significant optimisation potential through smarter routing alone.
Gartner currently places generative AI in what it calls the trough of disillusionment and forecasts that 25 percent of planned 2026 AI budgets will slip into 2027 as proofs of concept fail to survive procurement scrutiny. A separate Gartner study found that only 28 percent of AI infrastructure projects fully deliver against their original business case.
By the end of 2027, Gartner also expects that more than 40 percent of agentic AI projects will be put on hold as costs rise faster than business value becomes clear. That is not the technology going through an awkward adolescence. That is a market actively repricing itself.
None of this means AI coding tools will stop being used. The productivity evidence is real. At Uber, 70 percent of committed code now involves AI. At Microsoft, the engineering culture has already been shaped by these tools in ways that will not reverse simply because one vendor licence was cancelled. The demand is structural and it is not going away.
What is changing is the terms on which that demand gets funded. The era of uncapped AI experimentation on the corporate credit card is ending. The era of AI coding as managed infrastructure, with real cost governance, spending limits, model routing logic and ROI reporting, is just beginning. The companies that navigate that transition well will capture the productivity benefits. The ones that do not will find themselves explaining a very large and very unexpected bill.
Productivity may be real, but so are the bills. Treat AI coding tools like cloud infrastructure, not like a fixed-cost SaaS seat, or the invoice will arrive before the governance does.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
