Uber's CTO said it plainly: "I'm back to the drawing board, because the budget I thought I would need is blown away already." Claude Code adoption jumped from 32% to 84% of Uber's 5,000-engineer org between December 2025 and March 2026. By April, the entire annual AI budget was gone. Monthly API costs per engineer were running between $500 and $2,000. The per-token price during that period kept falling. The bill kept rising. This is the central confusion in how teams budget for AI at work, and most teams are still getting it wrong.
Why falling token prices do not mean falling bills
In late 2022, running a GPT-4-class model cost approximately $20 per million tokens. In early 2026, equivalent performance costs $0.40 per million tokens - a roughly 1,000x reduction in just over three years, one of the fastest cost declines in computing history. That number is real. It just does not tell you what you will pay.
Your AI token costs dropped 280x in two years. Your AI bill went up 320%. The gap lives entirely in consumption volume. A simple chatbot query triggers one inference call, but an agentic workflow - where the model calls external tools, verifies outputs, and self-corrects - can trigger 10 to 20 model calls for a single user-initiated task. Gartner's March 2026 analysis found that agentic models require between 5 and 30 times more tokens per task than a standard chatbot. Enterprises that scaled past the pilot phase discovered this multiplier only after their production bills arrived.
The pilot economics never reflected the production reality. A demo runs one tidy task. A production agent runs thousands of tasks per day, loops on failures, re-sends conversation history on every turn, and hands context packages between sub-agents. If Agent A sends a 5,000-token context package to Agent B, and Agent B processes it and sends a summary to Agent C, you are paying for those tokens multiple times across multiple model calls. This compounds further when agents loop. A planning agent might call a sub-agent five times while iterating on a plan.
The 87% gap hiding in your model routing decision
Here is the number most coverage of inference pricing skips. A Q1 2026 analysis of 2.4 billion enterprise API calls found that organizations running a tiered model architecture achieved a median blended cost of $2.31 per million tokens. Organizations routing every workload to frontier models paid $18.40 per million tokens. That 87% gap is the direct financial consequence of one architectural decision made at the start of the deployment process, and in most cases never explicitly revisited.
Frontier AI is priced for frontier tasks: complex multi-step reasoning, long-context synthesis, judgment under genuine ambiguity. Classification, extraction, intent detection, document summarization, and the routine logic that makes up the majority of most enterprise agentic workflows do not require frontier capability. But most teams deploy one model - usually the biggest one they trialed - and never route anything away from it.
The fix is unglamorous: build a routing layer before you scale. Send classification tasks and intent detection to a smaller model. Reserve the frontier model for genuinely ambiguous, high-stakes decisions. A cascade router can save up to 98% of a frontier model's inference cost while matching its performance on tested tasks. The figures depend on the model pair and query mix, but the pattern repeats across studies: a lot of traffic does not need the big model.
One more tax that compiles quietly: multi-turn context bloat. In multi-turn conversations, each subsequent API call includes the full conversation history. By turn 10, cost per call is roughly 7-10x the cost of turn 1 for identical output. An agent running a 15-step task re-sends its growing history on every step. Most teams never measure this.
Beagle in action#ops-team, Thursday afternoonThe steelman for not worrying about this yet
It is fair to push back. Inference costs have declined 30-50% annually for open-source models since 2023, and that trend continues as hardware improves and providers optimize.
Improved inference engines like vLLM 0.8.x deliver 30-50% higher throughput than versions from a year ago on identical hardware, meaning your infrastructure produces more tokens per month without any hardware upgrade. The cost curve is genuinely friendly.
The optimist case: keep shipping agents, the economics will catch up. Token deflation plus smarter MoE architectures plus hardware improvements means that today's alarming bills will look manageable in 18 months.
That argument is not wrong about the direction. It is wrong about the timeline. Teams have reported running three times over their annual token budgets by spring. One widely-cited incident saw a handful of agents stuck in a recursive loop run up tens of thousands of dollars in a single weekend. The discipline the industry built for cloud spend - FinOps - is now being rewritten for tokens, and most organizations admit they do not yet have the granularity to govern it. Waiting for cheaper tokens to fix an architecture problem is the same logic as waiting for cheaper cloud instances to fix a runaway Lambda function. The price drops; the bad code is still running.
Research on token allocation across agentic tasks found that token usage is highly variable and stochastic - some runs use up to 30x more tokens than others on the same task. Crucially, more tokens do not translate into higher accuracy: accuracy often peaks at intermediate cost and degrades at the highest cost levels, suggesting excess token expenditure frequently reflects unproductive exploration rather than deeper reasoning. You are not always getting what you are paying for.
What the per-seat pricing collapse means for your budget
There is a second economic shift happening in parallel that makes this harder to ignore. Pure per-seat pricing is shrinking: seat-based models fell from 21% to 15% of SaaS companies in 12 months. The reason is that AI agents make seat-based pricing structurally absurd.
AI agents are pushing software companies away from the per-seat SaaS pricing model because agents do the work themselves instead of giving a human a faster tool. When software completes a task end to end, charging by the number of logins stops making sense.
The replacement - outcome-based pricing - sounds cleaner. Intercom's Fin AI Agent charges $0.99 per resolved customer support conversation. If Fin does not fully resolve the ticket, the customer does not pay. Zendesk launched outcome-based pricing for its AI agents at $1.50 per automated resolution on committed volume, $2.00 on pay-as-you-go.
The trap here is that outcome-based pricing on the vendor side does not automatically control your inference spend on the infrastructure side. You can be paying $0.99 per resolved ticket to Intercom while simultaneously running a multi-agent loop inside your own stack that costs you $4 in raw inference per resolution. The vendor's pricing model and your infrastructure cost model are separate problems. Both require attention.
The point is not to chase the lowest cost per token. The point is to hit your product's quality bar at a sustainable unit cost. Track cost per successful outcome, not cost per million tokens. Those two numbers can move in opposite directions - and usually do.
AI agent inference costs: common questions
Why are my AI costs rising even though token prices keep falling?
For two years the story about AI costs was reassuring: per-token prices kept falling, so the bill would take care of itself. In 2026 that story fell apart. Per-token prices did keep dropping - but consumption exploded, because agents do not make one call, they make hundreds, in loops, across tools, unattended. The result is a new line item that lands on the CFO's desk with no warning.
What is the actual token multiplier for an agentic workflow vs. a chatbot?
Agentic workflows use 5-30x more tokens per task than a simple chatbot, according to Gartner's March 2026 analysis. A simple chatbot query triggers one inference call, but an agentic workflow - where the model calls external tools, verifies outputs, and self-corrects - can trigger 10 to 20 model calls for a single user-initiated task. The exact multiplier depends on loop depth, context size, and how often sub-agents hand off to each other.
Does routing to smaller models hurt output quality?
Not for most tasks. Frontier AI is priced for frontier tasks: complex multi-step reasoning, long-context synthesis, judgment under genuine ambiguity. Classification, extraction, intent detection, and document summarization do not require frontier capability. Route those to a smaller model and reserve the frontier tier for decisions that genuinely need it. Quality holds; cost drops.
How do I measure the real cost of an agentic workflow?
Track cost per successful outcome. Include model calls, retries, retrieval, evaluation, human review, monitoring, and incident handling. Cost per token is a billing metric. Cost per successful outcome is a business metric. Most teams have the first one and none of the second.
Is the per-seat SaaS model actually going away?
It is contracting, not collapsing. Pure per-seat pricing fell from 21% to 15% of SaaS companies in 12 months. But the shift matters most for tools whose entire value proposition is autonomous task completion - where a single agent replaces multiple human seats and the vendor cannot grow revenue by selling more licenses. Tools that amplify human work, rather than replace it, still price by seat without contradiction.