The price of AI is falling, but enterprise AI bills are climbing. The reason is a surge in data consumption as autonomous workflows, coding tools, and automation systems turn every task into a larger infrastructure event.
The price per million tokens has plunged over 90% across the industry’s leading models over the past two years. Yet lower model prices are not translating into lower AI bills for many organizations, because the volume of data moving through those models is undergoing exponential expansion.
Cheaper Models, Bigger Enterprise Bills
Much of the conversation around enterprise AI economics has focused on declining LLM inference costs. Enterprise CFOs and FinOps teams are seeing a different picture: total generative AI budgets are rising even as model prices fall.
Tools designed to act on behalf of developers or automation systems trigger an operational shift. The LLM context window becomes an unmanaged, highly variable layer of cloud infrastructure, and companies can end up paying a premium to transport low-signal, repetitive telemetry over external API channels.
A single automated troubleshooting session can rack up significant infrastructure costs by re-reading identical codebase metadata repeatedly. Agentic workloads add another layer of complexity because one task can fan out into dozens of model calls, with costs evaluated at execution boundaries instead of through one simple prompt.
Uber’s experience shows how fast spending can move. In December 2025, the company gave its engineers Claude Code and set up internal leaderboards tracking token consumption. By April 2026, Uber’s entire AI coding budget for 2026 was gone.
Uber caps spending at $1,500 per employee per agentic coding tool each month, but Uber President and COO Andrew Macdonald stated there was no link yet between overuse and shipping better products. Microsoft questioned the cost of Claude Code licenses before canceling them across its Experiences and Devices division, while Duolingo reversed a plan to factor AI use into performance reviews after employee pushback.
The scale of the market is moving in the same direction. Gartner expects spending on AI agent software to approach $207 billion in 2026, up more than 139% from $86.4 billion in 2025.
Stop Rationing Tokens, Start Fixing Defaults
Noe Ramos, VP of AI operations at Agiloft, argues that the problem is not simply employees using too much AI. “Teams aren’t burning spend because they love waste. They’re burning it because the default infrastructure pushes them toward it.”
Dmytro Palaniichuk, head of engineering at Promova, points to the settings behind that spending: “It turned out that on team and enterprise plans, the premium models are pre-selected; they run at high reasoning effort by default, and on our plan, Opus sessions upgrade to a 1M-token context window. Of course, nobody changes these in their own settings.”
Palaniichuk emphasizes conscious model selection and awareness of which model fits each task. High token usage does not automatically indicate waste, because an expensive model may reduce engineering work or prevent a costly mistake.
“The first reaction should not be to ‘use less AI’. High token usage can mean a lot of different things. If someone is using $16,000 in tokens to save us $100,000, of course we want to encourage that,” said Rick Spencer.
Companies are testing different ways to control those costs. SUSE sorts AI usage into buckets but does not automate routing; it plans to implement a proxy. SUSE has no metrics, though it reports anecdotal improvements, including a project that went from hundreds of CVEs to zero.
Everlaw runs per-person token caps that trigger a notification and usually double the cap for the next day. Its effort on core Java infrastructure consumed $3,500 in tokens and reduced engineering time from 9.5 to 2.5 engineer-months. “The ROI of $3500 to save seven months of engineering time is simply a no-brainer,” said Max Christoff, CTO of Everlaw.
Everlaw’s larger unlaunched product has run $27,000 in tokens and will likely reach $40,000, reducing engineering time from 90-100 to 19 engineer-months. Those figures show why a simple spending limit can miss the real question: what did the AI work deliver?
Smart Routing Becomes the New Cost Layer
Optimization is moving past negotiating cheaper bulk API contracts toward infrastructure management solutions such as smart routing. Agiloft scrapped token caps, made cheaper models the default, and routed tasks at the infrastructure layer rather than the prompt layer.
Many vendors, including Merge, Databricks, AWS Bedrock, and Azure AI Foundry, are addressing routing gaps with auto-routers. In June, Databricks introduced Smart Routing inside Unity AI Gateway at its Data + AI Summit, with beta features that include hard spend caps and cost attribution.
Databricks’ router evaluates requests using prompt intent, length, referenced files, stack traces, scope, reasoning depth, and execution complexity. It also considers that the same model performs differently depending on the harness used.
When routing reaches a budget ceiling, administrators can choose between a hard limit and a fallback to a cheaper model. Databricks observed teams routing routine tasks to lower-cost models without measurable drops in resolution rates, while its tools connect unified tracing, evaluation frameworks, datasets, trace analytics, and feedback loops to cost and quality monitoring.
Palaniichuk did not provide a before-and-after savings figure, but noted that Opus on the 1M-token window made up about a third of monthly spend. Validating accuracy for specific domains remains challenging because models vary and update cadences differ.
That leaves enterprise leaders with a more useful target than token scarcity. “Don’t ration the tool. Fix the architecture underneath it. Scarcity governance is a patch. Intelligent routing is the fix,” Ramos said.
The next phase of enterprise AI will not be measured by model prices alone. Companies will need infrastructure that matches task complexity to model cost, tracks the value created, and prevents repetitive data from turning every autonomous workflow into an open-ended bill.
Based on
