Why your AI bill tripled while token prices fell 75%

Enterprise AI bills are tripling even as token prices fell 75% over the past year because agentic workloads consume five to 30 times more tokens than chatbot tasks, according to an analysis by an unnamed engineering consultant. A 2026 audit of 30 production engineering teams found context repetition accounted for 62% of total inference costs, and Stanford Digital Economy Lab research shows agentic coding can be up to 1,000 times more token-hungry than simple code-chat, with costs varying by up to 30 times between runs.

Why your AI bill tripled while token prices fell 75% Token prices just hit an all-time low. So, why did your AI bill triple? I hear a version of that question in nearly every architecture review I run right now. The blended price per token has dropped roughly 75% over the past year, yet enterprise AI spending keeps climbing for any organization running agentic workloads. Finance is looking at the invoice, engineering is looking at the price sheet, and neither can reconcile the two. Agentic AI costs are fundamentally a control problem rather than a token pricing problem. As models grow more capable, they rely on extended reasoning loops — executing multiturn dialogues and multihop data retrievals to deliver meaningful results. However, every turn and hop expands the context window, consuming tokens with every iteration. Without strict path guidance and execution guardrails, an agent risks drifting down speculative paths, rapidly multiplying compute costs for marginal returns. The unit of cost changed, and the dashboard did not Both numbers are true at the same time. Per-token pricing fell. Bills tripled. Nothing about the pricing is broken. What changed is the workload being priced. A chatbot exchange is a single round trip. An agent handling a triage queue or a code review is a campaign. It plans, calls a tool, evaluates what came back, tries again when the result is thin, and escalates when confidence drops. Each of those turns is separately billable and a potentially new journey. Agentic workflows commonly run five to 30 times the tokens of a comparable chatbot task, which erases a 75% price cut and then some. Why agentic workflows burn so many tokens The largest hidden driver is context and lack of guidance. Every step an agent takes drags the accumulated history back through the model, so you pay for the same tokens on turn one, turn six, and turn 20. One 2026 audit spanning 30 production engineering teams put that repetition at 62% of the total inference bill.<sup 1</sup The other drivers compound on top of it: - Multistep reasoning: Agents do not think in one shot. They work through problems iteratively, and each iteration is billed. - Multiturn systems: Systems designed to retain context and state across back-and-forth interactions, allowing users to guide, refine, and build upon an AI's output incrementally through dialogue. - Multihop operations: Requires an agent to chain information together across multiple data sources, turning simple queries into complex reasoning loops. Every hop inflates the context window with new data, which means that without strict constraints, unguided hops will silently multiply your token spend. - Repetitive loops: A defined business or technical process is left for the agent to reason through. Outcomes are nondeterministic. - Tool calls: Every external tool interaction adds tokens on both the request and the response. - Context reinjection: State has to travel with the agent, so it gets resent constantly. - Retry loops: A failed step does not just waste time. It burns the same tokens a second time, and sometimes, a third. Research from the Stanford Digital Economy Lab puts agentic coding as much as 1,000 times more token-hungry than a simple code-chat exchange and found the cost of an identical task swinging by up to 30 times between runs<sup 2</sup — same task, same outcome, thirtyfold spread. The variance is the real problem because a number that moves that much cannot be forecast, only observed. Architecture makes it harder to see. Roughly 81% of enterprises run three or more AI models, and 88% use two or more providers.<sup 3</sup Cost data ends up scattered across separate billing consoles with no common unit of comparison, which means nobody in the room can answer a simple question: What does one completed task cost us? The enterprise impact goes well beyond the bill Poor cost visibility does not just produce surprise invoices. It kills projects. Only 23% of AI projects launched in the past year were fully deployed and meeting their original ROI objectives.<sup 4</sup When AI initiatives stall, the culprit is rarely model quality; it is the inability to articulate the relationship between the workload’s cost and its return. The instrumentation gap is stark. In a 2026 Elastic survey of enterprise IT organizations https://www.elastic.co/blog/2026-observability-trends-generative-ai-opentelemetry , 85% said they plan to implement large language model observability. Only 8% have actually done it. That leaves most teams running production agents with no per-task cost telemetry at all and working from vendor billing summaries that report aggregate credits and cannot attribute a dollar to a prompt, a tool call, or a retry. The downstream effect lands squarely on the AI engineer. If you cannot prove the ROI of the tokens you are spending today, https://www.elastic.co/blog/how-it-leaders-measure-roi-of-agentic-ai-initiatives you will not get budget approval for the next project. Leadership confidence erodes quietly, and the AI roadmap shrinks a quarter at a time. Start with what you can capture this week The teams that get out of this hole do not begin with a reference architecture. They begin by capturing five fields per agent run: tokens in, tokens out, model used, retry count, and whether the task actually completed. That is enough to compute a real cost per completed task, and it is usually enough to expose the two or three workflows quietly consuming most of the budget. Route the traffic through a single gateway so that those fields exist in one place across providers. Then, compare the top and bottom deciles of the same workflow. The gap between a cheap run and an expensive run of the identical task is where your optimization work lives, and you cannot see that gap in a monthly invoice. Partial telemetry beats perfect planning here. I have watched teams spend a quarter designing observability while their agents ran uninstrumented in production. The first week of real data usually reframes the roadmap anyway. Measure the task, not the token Agentic AI costs are rising not because pricing is broken but because the way we measure value has not caught up with the way agents work. Token spend was a reasonable proxy when one prompt produced one answer. It is a useless proxy when a single task triggers 40 inferences, six tool calls, and two retries. There are four metrics to replace it: cost per completed task, context efficiency, quality-adjusted efficiency, and business value multiple. Together, they shift the unit of measurement from the token to the outcome. And they give you something a token dashboard never could — answers to board-level questions like what is our AI actually costing us? Where is the value landing? Why did that initiative stall? Dive deeper into these agentic AI value metrics and how to improve them with our guide https://www.elastic.co/resources/article/measure-and-improve-agentic-ai-value-and-cost-it-leaders . 1. Sinch AB, “ Sinch research reveals 74% of enterprises have rolled back live AI customer communications agents https://www.prnewswire.com/news-releases/sinch-research-reveals-74-of-enterprises-have-rolled-back-live-ai-customer-communications-agents-302770730.html ,” 2026. 2. Stanford Digital Economy Lab, “ How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks https://digitaleconomy.stanford.edu/publication/how-do-ai-agents-spend-your-money-analyzing-and-predicting-token-consumption-in-agentic-coding-tasks/ ,” 2026. 3. a16z, “ Leaders, gainers and unexpected winners in the Enterprise AI arms race https://a16z.com/leaders-gainers-and-unexpected-winners-in-the-enterprise-ai-arms-race/ .” 2026. 4. Hyper Frame Research, “ State of the Enterprise AI Stack https://hyperframeresearch.com/hyperframe-lens-research/ ,” 2026. The release and timing of any features or functionality described in this post remain at Elastic's sole discretion. Any features or functionality not currently available may not be delivered on time or at all. In this blog post, we may have used or referred to third party generative AI tools, which are owned and operated by their respective owners. Elastic does not have any control over the third party tools and we have no responsibility or liability for their content, operation or use, nor for any loss or damage that may arise from your use of such tools. Please exercise caution when using AI tools with personal, sensitive or confidential information. Any data you submit may be used for AI training or other purposes. There is no guarantee that information you provide will be kept secure or confidential. You should familiarize yourself with the privacy practices and terms of use of any generative AI tools prior to use. Elastic, Elasticsearch, and associated marks are trademarks, logos or registered trademarks of elasticsearch B.V. in the United States and other countries. All other company and product names are trademarks, logos or registered trademarks of their respective owners.