AI Costs Explode as Token Consumption Proves Unpredictable and Usage-Based Billing Takes Over

What you'll pay for AI agents will be wildly variable and unpredictable

Users must demand price transparency and performance guarantees. Among all the challenges of implementing agentic artificial intelligence, the least-understood issue is cost. The providers of AI, such as OpenAI, Google, and Anthropic, have price lists, but none of those listed prices tell users what the final bill will be to actually solve a problem. The result, according to a new study of costs from the University of Michigan and collaborating institutions, could be sticker shock: soaring and unpredictable costs of agents. The study, by lead author Longju Bai of Michigan and collaborators at Stanford University, All Hands AI, Google's DeepMind unit, Microsoft, and MIT, titled "How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks," is, according to the authors, "the first systematic study on AI Agent token consumption." The study was posted on the arXiv pre-print server. It is noteworthy for having as its author a prominent Stanford economist who has commented extensively on AI's impact on productivity, Erik Brynjolfsson. The top-level finding is that agents consume orders of magnitude more tokens than turn-by-turn, simple, prompt-based chats -- think 3,500 times the number of tokens for an agent as for a round of prompts with ChatGPT. Also: AI agents are fast, loose, and out of control, MIT study finds A token is the fundamental unit of information processed by an AI model. It could be a piece of a word, a whole word, or just a punctuation mark, depending on how a model chops data into pieces. You might expect agents to cost more in tokens, but the study reveals more alarming facts. Two different models can have wildly different token costs for the same task. And the same model can have different costs each time that it works on the same problem, using as many as twice the number of tokens on one occasion compared to another. The worst part is that none of this can be predicted. Agents, Bai and team found, cannot reliably estimate how many tokens they will ultimately consume for a given task. "Agentic tasks are uniquely expensive," they wrote, while more tokens don't necessarily improve results. "Simply scaling token usage may not lead to higher execution performance," they wrote, and, "[AI] models systematically underestimate the tokens they need. The rising cost and the uncertainty of success are in no way accounted for in today's price lists from OpenAI and others. The work suggests there is no easy fix to the matter. The best users can do is to set hard limits on agentic computer use, possibly causing agents to halt before completing tasks. (Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.) The big picture is that users collectively will have to push back on OpenAI and the other vendors and demand some form of reliable cost estimation and guarantees of task performance. We reached out to OpenAI, Google, and Anthropic for comment. To study costs, Bai and team used the open-source agentic AI framework OpenHands, developed by scholars at the University of Illinois Urbana-Champaign and collaborating institutions. They used OpenHands to build agents, which they then tested on the open-source coding benchmark test SWE-Bench. The SWE-Bench tasks are taken from actual GitHub issues. Also: AI agents of chaos? New research shows how bots talking to bots can go sideways fast They first found the relative strengths of models. OpenAI's ChatGPT 5 and 5.2 "achieve strong accuracy at low cost," though they are not the most accurate. Anthropic's Claude Sonnet-4.5 achieved the highest accuracy but at higher token costs. Google's Gemini-3-Pro was somewhere in the middle. And the Kimi-K2 model from Chinese AI lab Moonshot may have the worst relative mix: the most tokens to achieve the lowest accuracy. The authors suggested the difference in tokens is based on unique properties of how models are architected: "The gap is not driven by task difficulty or by some models attempting harder problems. Instead, the same task is simply more expensive for some models than others, reflecting a behavioral tendency of the model rather than a property of the problem." But the issue is not one of better or worse models because even the same model can take twice as many tokens to solve the same problem from one "run" of the task to the next. "The most expensive runs double the token and monetary cost of the least expensive runs," they observed, "suggesting that the agent's token consumption has large variances even when working on exactly the same problem." The lesson is that more tokens don't necessarily get you better results. "Simply scaling token usage may not lead to higher execution performance," they wrote. In fact, the authors found that generally work can get worse the longer an agentic spends on a task. "Accuracy often peaks at intermediate cost and saturates at higher costs," they observed. "Agent behavior becomes increasingly unstable on more complex tasks." Many models seem to search and search to solve a problem even when it's fruitless. "Models lack a reliable mechanism to recognize when a task is unsolvable and stop early," wrote Bai and team. "Instead, they continue exploring, retrying, and re-reading context, accumulating cost without progress." Those factors make "token usage prediction and agent pricing a fundamentally challenging task," wrote Bai and team. And, in fact, the bot itself cannot predict when asked to "introspect," they found. Bai and team asked each AI agent to predict its tokens using the prompt: "I've uploaded a python code repository in the directory example repo. You are a TOKEN ESTIMATION agent. Estimate the token cost to fix the following issue description," and then the problem description, such as, fixing a bug for a comparison function in code that fails. What they found is that agents can approximate to a small degree how many tokens will be used, but their predictions tend to be too low "Models consistently underestimate the tokens they need," wrote Bai and team. "The bias is especially pronounced for input tokens, whose predictions stay compressed even as real values grow into the millions." That last point, about input tokens, has a special prominence in the report. Bai and team found that input tokens, such as what's typed by the human user, and what is retrieved via tools such as database searches, dominate the cost in tokens. The other two types of tokens, the output, which is generated, and the cached tokens held in memory from prior stages, are far less demanding. "Strikingly, input tokens, not output tokens, dominate the overall cost in agentic coding." The reason is that "agentic workflows accumulate the information from different sources and the same context gets fed into the models repeatedly." As a result, there is a "dramatically higher input/output ratio" for agentic AI than for single-prompt or multi-prompt AI sessions with a bot. And, drilling down even further, the most expensive input token factor is when the agent retrieves prior information from memory. "We find that cache reads dominate both raw token volume and dollar cost," Bai and team wrote. "In every phase, cache-read input tokens are the largest category by a wide margin (Figure 8a), reflecting the cumulative reuse of prior context." Overall, the study results confirm my anecdotal experience with coding agents such as Replit and Lovable, where the meter was constantly running to use the underlying AI models, and I had no sense of what the total cost would be. What can be done? The authors don't have many suggestions. One proposal is that even if agents can't predict the number of tokens, they can make some guesses at a high level, a "coarse-grained" estimate for token cost. "This suggests that agent-driven estimation can potentially support early budget alerts before launching expensive runs, improving cost transparency without overpromising precise token-level accuracy," they wrote. I can think of a few other sensible guidelines. Since input tokens are the biggest cost element, one should think carefully about what can be controlled at input. The size of prompts is one factor that drives input tokens higher. The context window used with an agent, wider or narrower, affects token count at input. And the number of tools called by the agent, such as databases, will bring lots more input tokens into play. Also: Can a newbie really vibe code an app? I tried Cursor and Replit to find out There's only so much you can do as a user, however. Something more will have to be done on an industry-wide basis. The problems outlined are clearly those of a young industry, and one where vendors will have to be pushed by users to change practices. The lack of transparency as to what an agent might cost to do a task is way too vague for enterprises that need to be able to plan investments in software. The burden is pushed onto the user to run agentic tasks in an experimental capacity over and over in order to get anything like an average cost to use as an estimate for planning purposes. And the lack of guarantees of success -- even after the agent burns through tokens -- is the most glaring problem. That means enterprises could waste vast amounts of money just running tokens. Users collectively are going to have to push back on vendors such as OpenAI, Google, and Anthropic and demand price transparency and some form of guarantee that a task will be completed, or else the entire exercise of agentic AI may be dominated by cost overruns and failed implementations. Such deep problems are probably already being encountered by early adopters. They may be content to pay such a high cost to be among the first to get an agentic edge. It's not a situation, however, that can lead to stable, steady use of agentic AI.

Locked, stocked, and losing budget: AI vendor lock-in bites

OPINION The days when you could jump from one frontier AI model to another at the drop of a hat are going away as vendor lock-in starts to kick in, and prices increase. Once upon a time, say last month, people thought nothing of jumping from one AI frontier model to another. One week, the hottest AI model was Gemini 3.1 Pro, then it was Claude 4.6, now, maybe, it's GPT-5.5. Next month? Who knows. That's fine for Joe Amateur Programmer, but for Janet Pro Programmer, it's another story. You see, enterprise AI buyers face two converging problems. First, it's proving much harder to switch between AI vendors than people expected. At the same time, AI vendors are pushing through price increases that are reshaping software economics. We always knew this would happen. AI prices have been loss leaders for years now, and the bills are finally coming due. A recent survey by AI orchestration platform provider Zapier of 542 US executives with active AI vendor contracts, found that nearly 90 percent believed they could switch AI vendors within four weeks, and 41 percent said they could do it in just 2-5 business days. Now who's hallucinating? I've long thought that behind all the lip service company brass gives AI, most senior executives are completely clueless about what AI is and how to deploy it. This kind of delusional thinking is proof. According to Zapier's report, only 42 percent of organizations that attempted to migrate between AI platforms report that it went smoothly. The remaining 58 percent? They say the process either failed outright or required significantly more effort than expected. Really. Who'd have thought it? The trouble stems from all the layers of technical dependency that early adopters underestimated. AI implementations require vendor-specific APIs, proprietary training data, custom tooling for model deployment, and deep integrations into existing workflows, none of which transfer cleanly between providers. According to Zapier: "The problem is that when AI is already woven into internal processes, connected to other systems, and tuned to specific workflows, it has dependencies, edge cases, and little adaptations that nobody documented because they were 'temporary.'" It's not just the software which is making it harder to move. As AI consultant Haroon Choudery pointed out: "Switching model vendors is no longer just an API migration. It is context, workflows, and institutional memory." Moving any of that from one vendor's platform to another isn't easy, and it's even worse if you don't have a handle on what you've got locked into those three areas. Guess what? Choudery observed, "Most operators I talk to haven't mapped any of them." I'm not surprised. This is yet more proof that your C-level executives don't have a clue about what they're doing by pouring their resources into AI as fast as they can. Some people I've spoken to seem to think that, because AI costs so little, even if moving from one to another is expensive, they can to afford it because the models themselves are so cheap. Yet AI providers which are losing money hand over fist are finally raising prices across the board. For example, OpenAI increased the cost for developers using its flagship GPT-5.2 model from $1.25 per input token in the previous GPT-5.1 to $5.75. Ouch! It's not just OpenAI. Anthropic confirmed a de facto price increase for its Claude enterprise edition on April 15, 2026, when it moved from fixed pricing to a dynamic usage-based model. Experts think this could double or triple costs for heavy-duty users. You don't have to be a hardcore AI developer to see this. For example, when I wrote this, you could no longer get a new GitHub Copilot subscription. GitHub is also restricting the compute you'll get from its individual subscription plans, while dropping access to Opus models entirely. I do hope you weren't planning on launching your business around GitHub Copilot. It's not just pure AI programs where you'll see this. AI costs are also pushing up prices for programs such as Microsoft 365. Everyone is gonig to do this. There will still be sweetener fixed-price tiers, but you'll get less compute power in them. Like it or lump it, we're heading to a token-based pricing structure and the end of fixed-price tiers. As Nick Turley, an OpenAI executive, said recently, "There's no world in which pricing doesn't significantly evolve." You think? As for those all-you-can-eat plans? Forget about them. They're history. These pricing changes reflect fundamental realities of infrastructure. Memory chip prices, in case you haven't noticed, are giving gold a run for its money. All those gigawatt AI data centers aren't going to pay for themselves either. As Datos Insights CEO and co-founder Eli Goodman told Reworked last year: "The most common myth is that AI works like regular software. That's not true; every query has a real cost. The provider's bill goes up when you use more." AI is not like Software-as-a-Service (SaaS), where costs shrink with scale. We talk about how much AI training costs, but every query you make and agent you launch costs you inference tokens. In short, the more you use AI, under its new pricing structure, the more it's going to cost you. Nik Kale, Cisco principal engineer and product architect, added: "Microsoft's increases aren't a temporary spike -- they're the beginning of a new price baseline for the AI era. GPU capacity, inference scaling, and the rising energy demands of large-model workloads have become structural, recurring costs. Vendors can't absorb them anymore." Can you? Well, you're going to find out. But, wait, there's more! Say you're running Meta Llama on your own hardware. You're safe then, right? Right? Wrong. First, Llama was never, ever really open source. So, when Meta decided to turn it into abandonware in favor of its proprietary Muse Spark, you're left in the lurch. "The question isn't whether AI is useful," the Zapier report noted. "What happens when the AI you depend on disappears, spikes its prices, or gets acquired by a private equity firm that's going to strip it for parts?" That's a darn good question, isn't it? Do you have an answer? You'd better start working on one. The more you've already invested in AI, the more you're almost certainly locked into specific vendors, and I guarantee you their prices are going to increase to everything the market can bear and then some more. ®

The Economics of Using AI to Churn Out Code Are Looking Worse Than Ever

Can't-miss innovations from the bleeding edge of science and tech In theory, using an AI model to churn out a bunch of code for your software company sounds like an awesome idea. A ruthless boss can cut down their workforce and save on paying their salaries and healthcare. Or they can keep the employees, and force them to crank out even more code with an AI. Either way, it's an efficient cost-cutting dream. But reality is beginning to set in. Businesses are racking up huge AI usage bills they didn't expect, with a single employee spending over $150,000 a month on AI tokens. And AI companies are feeling the squeeze that the rampant use of their coding tools is putting on their servers, causing them to jack up usage rates. In other words, the economics of rapidly deploying AI coding tools across companies are looking more questionable than ever. And one telling weather vane of where things are headed comes from Anthropic, which quietly doubled its estimate for how much the average business will end up spending on its Claude Code tool. The sneaky update was first spotted by Ed Zitron, who extensively covers the AI industry. Before April 16, the Claude Code document estimated the average cost per developer to be $6 per day, and the average cost for 90 percent of users to be below $12. But the documents now state that the "average cost is around $13 per developer per active day and $150-250 per developer per month, with costs remaining below $30 per active day for 90 percent of users." A difference of several dollars doesn't sound like much in isolation. But costs are clearly trending upwards, and those dollars add up. Many developers run multiple AI agents per employee at the same time, using them to churn out code for different tasks for hours on end. For organizations with thousands of employees all running their own posses of AI servants, the costs could be exorbitant -- are exorbitant, in fact. Some, by their own admission, could be paying human salaries for what they're spending on coding tools. Bryan Catanzaro, vice president of applied deep learning at Nvidia, recently told Axios that for his team, the "cost of compute is far beyond the costs of the employees." These questions are swirling amid AI companies doing some severe belt-tightening, cutting off free trials and limiting access to their coding models to even paid users, as Anthropic experimented with doing. This week, Microsoft's GitHub Copilot said it would transition to usage-based billing, effectively forcing users to pay more for the code they crank out. Meanwhile, a growing body of research has poked holes in the premise that integrating AI leads to actual productivity gains, such as an MIT study that found that the overwhelming majority of companies saw zero growth in revenue after adopting AI. Another study found that was creating a new office paradigm of "workslop," in which AI appears to be generating more work that in reality needs to be fixed by employees down the line, both bogging down workflows and breeding resentment. Yet more studies have shown that AI is actually intensifying work for employees and driving them towards burnout. As the costs of using AI rise, in other words, the question of whether AI is worth all the baggage it brings will be harder and harder to ignore.

Dark Cloud Gathers as Bill Comes Due for AI Industry

Can't-miss innovations from the bleeding edge of science and tech AI companies have swept customers onto the bandwagon by offering cheap and even free access to their AI models. But the bill is finally coming due, and the results could be ugly. Earlier this month, Microsoft's GitHub Copilot customers were informed that they would have to reduce their usage of the AI coding tool because of the "significant strain" placed on company servers. It also nixed free trials for new accounts, citing abuse of the system. Now, the popular development platform is doubling down on its cost-cutting regime. On Monday, it announced that all GitHub Copilot plans will be moved to usage-based billing, charging customers based on the number of tokens their AI tasks consume. GitHub said the change, which will go into effect on June 1, "aligns Copilot pricing with actual usage and is an important step toward a sustainable, reliable Copilot business and experience for all users." The company is replacing its previous "premium request units" with "GitHub AI Credits." With the old system, users could carry out a fixed number of these intensive "premium" requests. Some requests required more units, but the actual real-world cost of performing them wasn't factored in. With the new system, the AI credits are directly tied to the number of tokens a request consumes. Subscribers will get an allotment of AI credits equal to the dollar amount they're paying each month -- so a user on a $10 per month plan gets $10 in monthly AI credits -- and will need to pay for additional credits. In the announcement, GitHub Product chief product officer Mario Rodriguez called the old model "no longer sustainable." "A quick chat question and a multi-hour autonomous coding session can cost the user the same amount," Rodriguez continued. "GitHub has absorbed much of the escalating inference cost behind that usage." The remarks are revealing. Until recently, AI companies have been happy to shoulder the costs of providing relatively cheap access to their enormously power hungry models, bringing in hundreds of millions of users into their ecosystems. But with the rise of AI agents and coding tools, the tasks that AIs are being asked to perform are more demanding than ever. Companies are feverishly deploying them across their workforces, encouraging employees to use the tech as much as possible. Software engineers, happy to oblige, will run multiple AI agents in the background at the same time, which can churn out code for hours on end -- but accumulating significant costs on the backend. Microsoft's GitHub isn't the only company heading down this route. Anthropic has repeatedly tinkered with Claude code rate limits, and recently began imposing tighter limits during peak hours. It also experimented with cutting off Claude Code access to its lowest-tier paid users, drawing alarm from users. Google imposed weekly limits on its AI coding environment Antigravity earlier this year. Given the financial realities of the space, it's not surprising to see AI companies start charging more. But it'll be interesting to see how customers and business, which are already racking up huge AI bills, react to the cost changes -- and what it all means for AI adoption at large.

AI's Free Ride Is Over and the Tab Is Big | PYMNTS.com

By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions. Microsoft's GitHub announced that all Copilot plans will shift to usage-based billing on June 1. Fixed request allowances are out, replaced by a credit balance that depletes based on actual use. Base prices hold: Copilot Pro stays at $10 a month and Business at $19 per user. But GitHub said the previous model was no longer sustainable as infrastructure costs climbed. Anthropic made the same call. PYMNTS reported the company began charging enterprise customers based on their AI consumption. Claude Enterprise customers now pay a flat $20 per user per month plus a variable charge tied to computing capacity used. Previously, those customers paid up to $200 per user monthly for a fixed usage allotment. Fredrik Filipsson, co-founder of Redress Compliance, a firm that helps businesses negotiate software licensing agreements, estimated the changes will double or even triple costs for heavy Claude Enterprise users, according to The Information. When a flat subscription covered everything, developers had no reason to hold back. They ran long sessions, tried different approaches and experimented freely. Now each extended session carries a cost, with more capable models running at higher rates. Reaction to GitHub's announcement was immediate. Users argued the change reduces value even where the sticker price holds. GitHub's FAQ even included the question: "This just wiped GitHub's value moat -- why should I stay?" The company's answer: usage-based billing aligns costs more closely to actual value and gives developers freedom to choose which models they use. CNBC reported that Anthropic's Claude Code surpassed $2.5 billion in annualized revenue by February, up more than 100% since the start of the year. In response, OpenAI in April rolled out a new $100-per-month Codex plan targeting the same developer audience. Running powerful AI models at that scale costs money, and neither company is willing to offer unlimited access indefinitely. Traditional software costs tracked headcount. AI costs track activity. A single employee can generate thousands of AI interactions in a day. Another may trigger none. An automated process can run continuously without anyone watching the bill. PYMNTS reported enterprise AI invoices now resemble utility bills more than software subscriptions. Charges are tied to model activity, not employee count. Finance teams built around stable annual renewals now manage a cost structure with no prior reference point. The costs compound further down. According to PYMNTS, for every dollar spent on AI models, businesses spend $5 to $10 on integration, compliance and monitoring. The subscription was only the visible line item. GitHub is introducing admin controls that let organizations cap spending at the company, team or individual level. Anthropic's enterprise changes apply to accounts with more than 150 users. Both give procurement teams a mechanism for managing spend. Predicting it in advance remains a separate problem. PYMNTS Intelligence found that more than 8 in 10 CFOs at large companies are using AI or actively considering it. AI pricing models continue to evolve as adoption scales. The pricing pressure has a structural cause. Building and running frontier AI models requires enormous amounts of computing infrastructure. That cost compounds as usage rises. Model makers are not yet profitable at scale and usage-based pricing is one mechanism for closing that gap as adoption grows.