Why Is Your AI Bill So High When Tokens Keep Getting Cheaper?

Uber's 2026 AI coding budget was gone by April. Not squeezed. Gone - four months into a twelve-month plan, after roughly 5,000 engineers adopted Claude Code faster than the finance team had modeled. The company's CTO confirmed the overrun, and by June Uber had instituted a $1,500-per-month cap per employee, per tool, trackable on an internal dashboard.

The strange part: per-token prices were falling the whole time.

This is the defining AI finance puzzle of 2026, and it has a specific mechanical answer.

Why cheap tokens don't mean a cheap bill

The answer first, then the mechanics: token price and token volume are independent variables, and agentic adoption has grown volume faster than price has fallen.

In late 2022, running a GPT-4-class model cost roughly $20 per million tokens. By early 2026, equivalent performance costs $0.40 per million tokens or less. That is a real and remarkable collapse. The price of a fixed level of AI performance has fallen roughly 1,000x in three years, per analysis from a16z.

But the workloads changed underneath the prices. The models that drove down cost-per-token are also the models capable of running multi-step agentic tasks - and those tasks consume tokens in a fundamentally different pattern than a chat window does.

Unlike standard generative AI, where tokens are used in a single linear transaction, agentic workflows use tokens recursively to reason, use tools, and self-correct. Gartner estimates agentic models require 5 to 30 times more tokens per task than a standard GenAI chatbot.

That 5-30x figure is for well-behaved agents. For reasoning models doing extended thinking, the multiplier is worse.

Reasoning models - OpenAI o1/o3, Anthropic extended thinking, Gemini thinking - bill for internal reasoning tokens not returned in visible output; these can be 5-50x the visible volume and dominate cost. You send a prompt, see a three-sentence answer, and get billed for the model's entire internal monologue. A paper measuring real agentic workloads found that Gemma Thinking spends 45.8-67.6% of its output on thinking tokens across workloads, while Qwen Thinking spends 29.0-40.7%. Enabling reasoning changes not only total token volume, but also the semantic composition of generated tokens.

The compounding effect: every tool call injects a full tool description into the model's context. Every reasoning step carries the growing conversation history forward. In production environments, this cumulative expansion of the conversation history often becomes the single largest hidden cost of an agentic workflow.

What Uber actually got wrong (and it wasn't the budget number)

The easy takeaway from the Uber story is that AI is expensive. That misses the actual lesson.

Uber's total research and development spend reached $3.4 billion in 2025, up 9% year over year, which makes the budget collapse less about scale and more about a pricing model that enterprise finance teams have not learned to manage.

The structural problem was a mismatch between the billing model and the budget model. Claude Code does not price on a per-seat basis - it meters tokens consumed across model calls. An engineer running autocomplete suggestions consumes a fraction of what an engineer orchestrating parallel agents across a monorepo will consume. The same tool, the same engineer, the same workday, can produce wildly different invoices depending on workflow choice.

Uber then made it worse organizationally. The company ranked engineers on internal leaderboards based on Claude Code usage. That created a cultural incentive to consume more tokens, which translated directly into faster budget burn. The teams driving adoption were not the same teams managing the spend, and that gap turned out to be the load-bearing flaw.

Uber burned through its entire 2026 AI coding budget by April after roughly 5,000 engineers pushed token consumption beyond projections. Walmart, Amazon, and Cisco followed with similar controls.

The math teams are not doing

Here is the comparison most coverage skips. If you model AI tooling as a per-seat SaaS product, you budget a fixed number. If you model it as a consumption product, you need a distribution - not a single number, but a range based on how your team actually works.

The per-token price of a frontier API call is a poor predictor of your monthly bill. The actual cost drivers are:

• Workflow type. A single-turn summary call might use 2,000 tokens. An agent resolving a multi-file refactor might use 200,000 - the same model, one click.
• Reasoning mode. Turning on extended thinking on a frontier model can multiply the token bill by 5-50x per call, as the model bills for its internal chain-of-thought.
• Context accumulation. Multi-step pipelines carry prior context forward with every step. A ten-step agent workflow doesn't cost 10x one step - it costs more, because each step's context includes all the previous ones.
• Retry behavior. Agents that self-correct on failure can loop. Without a hard token budget ceiling, a failing agent is a billing runaway.

Reasoning models can consume 100x more tokens internally than they output, creating a cost paradox where cheaper per-token pricing leads to higher total bills.

The number that matters is not cost per million tokens. It is cost per successful workflow completion. Once repeated context can be reused, the expensive part of a request is no longer all those tokens - it is the fresh, uncached portion and the output path. That pushes teams toward a better unit: dollars per successful workflow step, not dollars per raw token count.

A practical routing split looks like this:

For many workloads, per-token cost has dropped by an order of magnitude. The teams that win are not simply waiting for cheaper model pricing - they route routine work to smaller models, cache repeated requests, control context size, batch offline jobs, and measure cost per successful outcome instead of cost per token alone.

The structural shift that actually matters

Gartner's forecast implies commodity LLM inference will approach near-zero cost within 3-4 years. The strategic implication: the competitive moat in AI agent systems will not be access to cheap inference, but rather the quality of agent architecture, memory systems, tool integrations, and organizational knowledge embedded in agent behavior. Infrastructure cost will cease to be a differentiator; the quality of what agents do with that compute will be the remaining axis of competition.

That is the real second-order consequence most teams are not thinking about. The inference cost problem is solvable and will largely solve itself as prices keep falling. The harder problem - building agent workflows that actually complete tasks reliably rather than spinning in reasoning loops - is where the organizational work sits.

The FinOps Foundation's 2026 State of FinOps Report identifies AI and data platforms as the fastest-growing new category of enterprise spend, with token-based pricing, agent step billing, and retrieval costs introducing dimensions of cost volatility that legacy budgeting frameworks cannot handle.

Teams that build good measurement now - cost per workflow step, not cost per token - will have an advantage when the hard question arrives: which agent tasks are worth running at scale, and which ones aren't.

A teammate like Beagle, which operates on a draft-and-approve model inside Slack, gets a structural benefit here: a human stays in the loop on every send, which naturally bounds runaway loops. The cost accounting problem doesn't vanish, but you get a natural audit point.

Why is my AI bill so high: common questions

Why do AI bills keep rising when token prices are falling?

Token volume is growing faster than token prices are falling. Agentic workflows consume 5-30x more tokens per task than simple chat, reasoning models bill for internal thinking tokens that never appear in the output, and context accumulates across multi-step pipelines. Lower per-token prices do not offset those multipliers if teams are running more complex workflows.

What is the "agent tax" on AI inference costs?

The agent tax is the gap between what a task costs as a single API call versus what it costs wrapped in an agentic loop. Reasoning steps, tool-call descriptions, context re-injection at each step, and retry behavior stack up. One analysis puts the typical multiplier at around 50x for complex agentic tasks versus a simple prompt-and-response call.

How should teams budget for AI coding tools like Claude Code or Cursor?

Model three workflow types separately: quick autocomplete calls (low token count, predictable), single-file edits (moderate), and multi-file agentic tasks (highly variable). Set per-employee monthly caps per tool, and track actual cost per task completed weekly - not cost per seat or cost per million tokens. Avoid leaderboards that incentivize raw usage.

When does it make sense to use a reasoning model versus a standard one?

Use reasoning models - models that bill for extended internal thinking - only where accuracy has high business value and where a standard model demonstrably fails. For summarization, drafting, and code autocomplete, a mid-tier model is almost always the right call. Routing discipline matters more than which frontier model you pick.

Is the AI inference cost problem going to solve itself?

Prices will keep falling - the trend is durable and hardware improvements continue to push cost-per-token down. But total bills will keep rising in parallel as adoption deepens and workloads get more complex. The teams that come out ahead will be those that measure what tasks actually cost to complete, not those waiting for tokens to become cheap enough that measurement doesn't matter.