OpenAI weighs token price cuts as businesses confront high AI costs and switch models

AI Price Wars Begin: OpenAI Considers "Drastic Price Cuts" In Pursuit Of Anthropic Customers Earlier today, in a report discussing how "AI bills are out of control", JPMorgan tech guru and TMT salesman, Mark Schilsky wrote that "most of my high level investor discussions focus on one major topic: when will the party end? Put another way, tech investors have made so much money in Semis so quickly that they are looking for potential warning signs that the music is about to stop. Predicting such an end is incredibly difficult. As such, investors are searching for forward-looking indicators that might suggest the AI party is nearing a peak."  Here, the JPM trader highlighted perhaps the clearest indicator that the music was about to stop: "A slowdown in the growth of the annualized run-rate revenues of the major AI labs. If there is any sort of second derivative ‘kink’ in their growth algorithms, that could portend a future problem for the AI trade." In response to this, we pointed to just such a "slowdown in the run-rate revenues", when we showed that the Silicon Data token price index is down for 7 straight days to a level last seen in mid-January, or long before the current agentic craze started. Almost as if it knew something...  Source Turns out it did: late on Wednesday, with futures surging and Korean stocks erasing a nearly 5% drop and turning green, and euphoria generally back front and center, the WSJ may have burst the AI bubble when it reported that - contrary to conventional wisdom that token prices will magically go to infinity - OpenAI, which has been badly lagging both the revenue and IPO race with Anthropic in recent months - was considering "drastically lowering the prices it charges users" in a panic scramble to regain market share and win back customers from archrival Anthropic. And so, at a time when there is suddenly a mass realization that token prices had been soared in recent weeks, a wake-up call which JPM lovingly described as follows: "investors have been discussing the possibility that much of the token spend that corporate America is currently incurring is ‘wasted’. Anecdotes from companies like UBER aren’t helping this narrative", OpenAI is weighing significant cuts to what it charges for tokens. Hilariously, the move would be in anticipation of similar cuts the company expects at Anthropic, which is trying to double how much it charges for its latest model, Fable, which provides at best a very modest modest improvement in performance over Opus 4.8. In short, we now have a classical deflationary race to the bottom, precisely the opposite of what the profit-strapped industry desperately needs to grow into its gargantuan balance sheets (and massive SPVs); Instead, the AI world is about to get hit with a collapse in both revenues and profit margins, while cash burn goes into full-on incinerator mode. Warning that "business executives have begun to balk at the high prices for AI usage", the WSJ writes that OpenAI CEO Altman said at a recent event that costs had become “a huge issue.” “I think we’ll have a lot of ways we can help people get more value for less spend,” he said. In other words, LLMs tried to push up token prices to and beyond their breaking point... and succeeded. And now it's time for the brutal drop: a drastic price war will erode the profit margins of both companies, which already lose billions of dollars because of the enormous cost for computing resources needed to run AI systems.  Altman's decision to start a price war was prompted by OpenAI's attempt to catch up with its younger rival in the race to win enterprise customers that are paying large amounts of money for AI tools that can improve workplace productivity. Anthropic’s revenue recently surged"after its coding tool Claude Code went viral among software engineers, and the five-year-old startup surpassed OpenAI’s valuation for the first time." Or at least that's the WSJ version of events. In reality what happened is that Anthropic quietly annualized the one-time bumper revenue from Feb-May during the agentic splurge when nobody had any idea what they were paying, to come up with the ludicrous $47BN ARR, which they then actively paraded ahead of their IPO. But let's see what Anthropic's ARR is next month will be after clients finally check their token bills. Sure enough, as we have been writing repeatedly in recent weeks, "some corporations poured so much money into Anthropic’s products that their leaders are now seeking to rein in spending. Earlier this year, an Uber executive said the company had maxed out its 2026 budget for agentic, or autonomous, AI use, and another company leader said last month that it was difficult to link AI coding productivity improvements to new customer features." In other words, yet again the age-old question of whether and when AI will have a positive ROI rears its ugly head, and the answer is not any time soon... if ever.  Such comments from many executives have triggered a broader debate within Silicon Valley about so-called “tokenmaxxing,” or the practice of using as many tokens as possible to boost productivity, including in ways that don’t generate returns on investment. That may have worked 6 months ago when LLMs were giving out compute for free to capture market share, but it doesn't work now that all the major AI companies are suddenly charging an arm and a kidney for an "agent" that responds to emails.  As the WSJ concludes, "a price war would be an early test of the strength of both companies’ business models ahead of hotly anticipated public listings." OpenAI and Anthropic have captured the majority of revenue from new AI products, powering their rise. But an underlying risk that investors have long identified is the interchangeability of their products, and the ease with which customers can abandon one for the other.  There is a bigger risk: as we noted one week ago, in the coming price war, neither OpenAI nor Anthropic will win. Instead it will be the country that has made reverse engineering Western technology and then selling it back to the west at 90% off, into an art form. Yes, China is about to enter the chatbot.  Tyler Durden Wed, 06/10/2026 - 23:03

"A Lot Of BS, Honestly": Apollo Head Says Everyone Is Measuring AI Wrong The tokenomics debate got its sharpest contrarian voice this morning, and it came from inside the building. John Zito Photographer: Jeenah Moon/Bloomberg John Zito, co-president of Apollo Asset Management, sat for a fireside chat at the Morgan Stanley US Financials Conference on Wednesday, where he suggested to Bloomberg that measured per unit of intelligence delivered rather than per token, prices are collapsing - even as low-value usage drives the actual bills up. "I think tokenmaxxing and token talk is - it's a lot of BS, honestly. Like, if you look at per unit of knowledge and cost per unit of knowledge, prices are collapsing. Prices are collapsing per unit of IQ, if you did it that way." In other words: a token is not a unit of intelligence. Price the capability instead of the throughput - the way a 2026 laptop costs what a 2010 laptop did but is 50x more capable - and the cost of "IQ" is in freefall even while the bills explode. He also suggested that we're screwed if AI isn't just hype: "If AI is real, it's so hyper-deflationary to so many things over the long term that it's really hard to take risk." So a few things are going on here - the spending problem is real. The metric everyone is using to describe it is wrong. And the resolution of that tension is where the entire AI trade goes next. As Goldman's Rich Privorotsky noted five days earlier - consensus was already migrating to exactly this frame; that the relevant economic metric is not token volume but useful task completion per watt and per dollar; that customers facing usage-based pricing "will optimize for cost per completed task," routing simple work to local models, harder tasks to the cloud, and frontier models only when required; and, that we "maybe have allocated too much spend to the Data-Centric model." When Apollo and Goldman independently land on the same framework inside a week, we're looking at a new institutional consensus forming in real time. The French toast economy Zito's diagnosis of why enterprise AI bills exploded will sound familiar to ZeroHedge premium subscribers: too many companies pointing frontier models at tasks that don't remotely justify the compute. "Our IQs are so low that we're actually using [AI tools] to check out the recipe for, you know, French toast. That's where you're seeing the prices go up." Swap "French toast" for "checking the weather" and that is, almost verbatim, the tokenmaxxing reductio we documented at Amazon - employees routing busywork through agents to climb the KiroRank leaderboard, frontier reasoning models deployed against questions a search bar answered in 2009. Zito even joked that his own IQ is "not high enough" to need what Anthropic's next flagship model - and something that only "a handful" of users genuinely need, and can monetize, the bleeding edge. So - the mismatch between task and tool doesn't persist forever - it gets arbitraged into what he called a new economy for the sector: "The AMD chip, the Nvidia chip, all these different chips will be used and optimized for a certain use-case to solve the spend problem." Citadel and Jane Street pay anything for the frontier because their ROI is, in his word, massive. Everyone else's French toast queries get routed to something cheap. And of course we watched this unfold over the last month. Bloomberg notes Uber set usage limits on tools like Claude Code after incinerating its AI budget, and Walmart capped an in-house AI agent - the one that helps employees with spreadsheets and presentations - after demand ran too hot. The caps are landing on exactly the low-IQ-task tier Zito is describing, while the frontier spend stays untouched. How we got here For readers just joining: this is the latest beat in a story that has moved very fast. Last month we noted that the AI narrative had hit a serious snag, after Uber's COO Andrew Macdonald admitted the company couldn't draw a line between exploding token consumption and useful product output. This, after 5,000 engineers burned the entire 2026 AI budget by April. Data spanning 2,444 companies suggested only 18 cents of every AI dollar reaches users as stable product, with 44 cents going to fixing bugs the AI itself introduced. h/t @Aiswarya_Sankar Then came the $500 million mystery bill - an unnamed enterprise client, per Axios, torching half a billion dollars on Claude in a single month with no usage caps - landing the same week Amazon nuked its internal leaderboard and an SVP begged staff not to use AI for the sake of using AI. Then, in Part II of our reporting: 'From Singularity To Tokenomics,' we noted that the subsidy formally ended: GitHub flipped Copilot to usage-based billing on June 1 - the same morning Anthropic confidentially filed its draft S-1 - and developers hit their monthly quotas before lunch. OpenAI, Google, and Microsoft all executed the same flat-rate-to-meter pivot within sixty days of each other. Sam Altman conceded that cost went from a non-issue in January to, in his words, a huge issue and a meme. By Monday, Goldman's one-delta desk was flagging that the Silicon Data Token Spending Index had started to soften - Q1 may have been peak token-maxxing-as-KPI - and Citrini Research had coined the inevitable sequel: in a matter of weeks, the narrative went from tokenmaxxing to tokenpanic.  We're happy everyone is now looking at this chart... You're welcome? The Silicon Data LLM Token Expenditure Index rolls over What enterprise customers are actually saying Fresh comments from UBS paint an interesting picture. After polling actual IT execs at enterprise AI customers, the bank reports that token costs have become a real issue for roughly 60% of the enterprises they spoke to - "this is not a made-up media story," in the bank's own words. One customer described the GitHub Copilot pricing change in a single word: "chaos." Another got their first AI bill and heard leadership say, flatly, "we don't have the money for this." A third admitted: "we overbuilt in certain areas and are starting to feel the wrath." Source: UBS Evidence Lab But bears should pay attention to this part: not a single check was slamming on the AI brakes. UBS found the dominant behavior is guardrails, not retreat: caps, alerts, model-downshifting, pooled tokens - normal enterprise cost-containment. Several customers explicitly refused to throttle usage ("we don't want to throttle them... our aim is to just get our employees to start using AI") and are instead cannibalizing other IT spend - cutting external IT services, consolidating cloud, and notably, metering headcount growth - to make room for the AI line item. Even Uber, the poster child for budget incineration, has set per-engineer token caps around $1,500 a month - which, as UBS dryly notes, is still extremely high - and its CEO describes the company as full steam ahead. So according to UBS, costs have been spiking because adoption is ramping, not because per-unit prices are inflating - the per-unit cost of intelligence is falling. Which is exactly Zito's "cost per unit of IQ" point. Both things are true So is tokenmaxxing "a lot of BS"? Zito is right about the denominator. Cost per unit of intelligence is collapsing, relentlessly. Open-source and Chinese models deliver near-frontier capability at 10-25x lower cost; Cursor's new model matches frontier coding performance at a tenth the price per task. Measured per unit of IQ, this is the most deflationary technology in living memory. Zito's denominator: inference costs collapse for a fixed level of intelligence The CFOs are right about the numerator. Gartner has found that even a 90% collapse in inference costs won't make enterprise AI cheaper, because agents devour tokens faster than prices drop and providers don't fully pass the savings through. It is also the lived experience of every company in the UBS checks. A collapsing unit cost times an exploding unit count is still a bigger bill. Token expenditure is a meaningless productivity metric and a decisive revenue metric. Nobody underwriting a near-trillion-dollar AI IPO can dismiss it as a measure of revenue durability. That is why the rollover in the Silicon Data index is worth watching: the chart measures nothing about value created, and everything about the thing the valuations are built on. The same logic applies to the narrative itself. Zito calls the token talk noise; that noise doubled the market value of the semiconductor industry in two months on the way up and is unwinding it now. A fundamental investor can dismiss what a narrative measures. A trader cannot dismiss what it moves. What the 'noise' moved: semiconductor market value doubled in two months on the tokenmaxxing narrative The unresolved question for traders: The infrastructure complex is priced for token demand going up and to the right at the frontier - Goldman's 24x by 2030. But if Zito's use-case economy arrives, a large share of that volume migrates to commodity inference: cheap chips, open models, local hardware. Volume can keep growing while the dollars - and the margins - pool somewhere other than where today's valuations assume. The UBS checks already show the mechanism in motion: enterprises aren't cutting AI, they're cutting around AI, and routing down-market wherever good enough will do. Meanwhile, the token expenditure index printed its sixth straight down day - the longest streak since January - with Citadel's read attributing the drop to adoption becoming "less about what frontier models can do and more about the price," a shift toward cheaper models. Note what that means for the chart: six red days on an expenditure index isn't necessarily usage falling - it may be the deflation itself arriving in the spend line, the same work bought cheaper. The numerator and the denominator, colliding in one print. And savor the garnish: Citadel is one of the two firms Zito named as gladly paying anything for the frontier - and it's their desk narrating everyone else trading down. Token prices down 6 days in a row: longest streak since January. "Adoption is becoming less about what frontier models can do and more about the price... the recent drop in the token index may reflect some of this shift toward cheaper models"- Citadel - zerohedge (@zerohedge) June 10, 2026 Volume and value have decoupled. The desks have noticed. The repricing is the part that comes next. Tyler Durden Wed, 06/10/2026 - 14:20

From Token-maxxing To Token-panic: Citrini Warns AI Goldilocks Narrative Hitting A Wall When the world and their pet rabbit was buying the hype and extrapolating trends to infinity and beyond, we dared to highlight a few 'economic' realities of the new 'tokenomics'. From Singularity To Tokenomics: The AI Narrative Just Hit A Serious Snag Was Amazon's Tokenmaxxing Fiasco Behind Claude's $500M Mystery Bill? From Singularity To Tokenomics, Part II: The Subsidy Just Ran Out - And GitHub Users Went Splat This morning we got confirmation of this AI reality questioning from none other than Goldman Sachs Partner, Rich Privorotsky, who highlighted that Token Spend had 'peaked'... And now, Citrini Research - who infamously issued a less than utopic view of the world under AI back in March - has written a follow up on the status quo of the AI ecosystem, noting that in just weeks we’ve gone from tokenmaxxing to tokenpanic. In March, we and many others were writing about the astounding growth in token consumption driven by the release of agents and more intensive models. This was enough to send the infrastructure trade sharply higher – the market value of the semiconductor industry doubled in two months. But that goldilocks narrative is beginning to hit a wall. The corollary of explosive token usage is explosive cost to customers, which is coming just as the US labs and hyperscalers are turning up the dial on monetization. The public story is increasingly turning to corporate pushback. The first real signs of this shift were from the much-discussed report of Uber burning through its entire AI budget in just four months. Then there was the anonymous report of a $500 million oopsie. In the past week, the idea has turned into a media avalanche. According to The Economist’s reporting, Anthropic’s ARR has increased 5x since the start of the year, reaching $45 billion in May. Great for the lab, but it also means the “AI Opex” line item on P&Ls is going through the roof. The issue is not just Anthropic. Sam Altman also confirmed that all of a sudden cost is a huge issue (and acknowledged the virality of the idea). “Probably the second biggest theme is just around cost. People are really saying, it’s kind of become a meme now, but, “My company spent my entire 2026 budget in Q1. Can you make this more efficient?” We are continuing to push on that more with models. I think we’ll have a lot of ways we can help people get more value for less spend, but that went from, at the beginning of this year, an issue that never came up. I know. People were totally happy with the amount they were spending, to all of a sudden, a huge issue.” Microsoft’s AI Chief added to the unflattery this week after cancelling Claude Code licenses in May. “Anthropic is extremely expensive, and I think many people are urgently looking for alternatives” This cost concern didn’t just come out of nowhere. First, agents and more advanced reasoning models use orders of magnitude greater tokens. Corporates have widely distributed these tools and encouraged their use just as the average user was gaining the ability to casually run enormous bills. Second, prices for frontier models are increasing as providers are flipping to usage models and preparing for public market debuts. In a unified front – OpenAI, Anthropic, Microsoft, and Google – have all implemented pricing shifts towards usage/tokens, as they simply can’t afford to endlessly subsidize their products for power users. April 2: OpenAI changed Codex pricing to align with API token usage instead of per-message pricing May 19: Google changed Gemini subscriptions from “daily prompt limits” to a “compute-used” model. June 1: Microsoft’s GitHub Copilot transitioned to usage based billing And what does a rate sheet mean really if you have no idea what your usage burns in practice? Claude’s Opus 4.7 & 4.8 have the same “list price” as prior versions, but use a “new tokenizer” that may use up to 35% more tokens for the same fixed text. Is this an existential problem or just the VC playbook at unimaginable scale? Subsidize demand, gain market share and lock-in, then monetize. After all, companies are spending a trillion in capex to make trillions in revenue, right? Well either way we’ve reached Monetization, and maybe not by choice. As fast as lab revenue is growing, the fundraising has grown even faster. The money going towards building and running AI has exploded. The deepest pockets in the world – hyperscaler cash flow, venture capital, sovereign wealth, public credit, private credit, public equity – are footing most of the bill. Eventually, customers have to start picking up the tab. Free-AI is ending. Tokenomics is beginning. What happens when underlying costs of compute become more transparent and directly traceable to outcomes? The ROI debate is about to be answered in real time, across millions of users and use cases. For the median user, maybe not a whole lot changes. But science projects, freewheeling agents, and curiosities will either get cut or offloaded to open source models. Companies will restrict AI functionality and invest in oversight and observability. Budget constraints will pit AI spend against headcounts. Providers will become more competitive on pricing and will begin to optimize physical and digital architecture for efficiencies. In many (most) situations, good enough will do. The cost of running open-source, discount, or mini models is going down while their capabilities only improve. This week saw another batch of open source models like Nvidia latest Nemotron family which includes advanced general-purpose models as well as highly efficient, compact versions optimized for local deployment and specialized agentic uses. As the frontier continues to advance, inference costs drop precipitously for a fixed level of intelligence. Why rent a Ferrari when a Vespa does the trick? Of course, frontier models with highly specialized functions can continue to command an intense premium, but will serve a smaller segment of the market. A top lawyer can still bill at thousands per hour, even if millions of other workers are making minimum wage. But even across the high end, the gap between US and Chinese offerings is worth noting. Qwen 3.7 and Deepseek V4 are still behind Opus 4.8 and GPT 5.5 in terms of benchmarks, but they are 10x - 25x cheaper. Since releasing V4 Pro and V4 Flash in April, Deepseek has shot past Anthropic to the top of the charts on OpenRouter in terms of tokens processed. Meanwhile, Cursor, one of the most used coding agents, released their new model that was post-trained on compute provided by xAI after their $10 billion deal. The base model is a different Chinese open source model by Moonshot and it was trained on data Cursor gets from its customers. The results are even stronger than Deepseek, it’s comparable to 4.7 and 5.5 for 10x lower cost per task and is one of the fastest frontier models. There are obvious other “considerations” for large US enterprises that may prevent a mass exodus to Chinese alternatives. Plus, greater integration into workflows adds to lock-in. But there is a growing trend of application layer companies that will continue to post-train on open source base models for specialized workflows like coding and legal. But what does this mean for the AI trade? First, to be clear, revenues for labs and hyperscalers are going to grow. Token usage for top Anthropic models continues to go higher. Regardless of the pushback, frontier models can certainly create meaningful value especially in high-stakes fields like tech and finance, and there are still plenty of levers to pull in the monetization phase. The entire point is for them to start making money. Likewise, this won’t fix near-term compute constraints. But we do think that cost and efficiency only become more important as the bills get bigger. Themes of local inference, miniaturization, smart routing, observability, price competition, and efficient model architecture will grow. Competitive pressures and price competition are likely to stay. Subscribers can read the rest of Citrini's note here... Tyler Durden Wed, 06/10/2026 - 10:20

Uber's CTO told the world this month that the company spent its entire 2026 AI allocation by April. The story has been reported in a handful of outlets, hit the front page of Hacker News for 397 points and 469 comments, and is mostly being read as a cost-of-AI-tools story. It is one. It is also, on a closer reading of the numbers, a pricing-model story — and the structural fact that almost none of the coverage has emphasized is the one that determines whether this is a one-company anomaly or the beginning of an industry-wide budgetary crisis. The structural fact is that Claude Code, like most enterprise AI tooling in 2026, is priced on token consumption, not per-seat licensing. Token-based pricing scales with how aggressively the tool is used. Per-seat enterprise SaaS pricing — the model corporate IT budgets are built around — scales with how many people have access to it. Those two cost curves diverge in exactly the territory where productivity tools are designed to operate: high-engagement, daily-use, gradually-deepening workflows. The Uber data is the first public-facing version of a math problem most enterprise IT departments are about to discover privately. The numbers Uber CTO Praveen Neppalli Naga, named in Yahoo Finance's and Benzinga's coverage, said publicly that Uber is "back to the drawing board" on AI budgeting after the surge in Claude Code use blew through internal projections. The specific numbers, as reported across the multiple outlets covering the story: Claude Code adoption inside Uber's ~5,000-engineer organization went from 32% to 84% over four months. 70% of committed code at Uber is now AI-originated. 11% of live backend updates are "being written by AI agents built primarily with Claude Code," per the reporting. Per-engineer monthly API costs: $500 to $2,000. Uber's annual R&D spend is around $3.4 billion, of which the AI tooling line was a much larger fraction than expected. Cursor adoption plateaued; Claude Code dominated. These are extraordinary numbers in two directions at once. The productivity numbers — 70% of committed code, 11% of live backend updates produced by Claude-Code-driven agents — are the kind of figures that explain why Uber will not, and almost certainly cannot, simply turn the tools off. The cost numbers — $500–$2,000 per engineer per month, multiplied across an aggressively-adopting 5,000-engineer organization, eating an entire annual budget in four months — are the kind of figures that explain why the CTO is publicly using "back to the drawing board" language. What sits between the two sets of numbers is the pricing model. How the curve actually behaves Per-seat SaaS pricing has a single dominant variable: headcount. You buy 5,000 seats; you pay 5,000 × $X. The cost is predictable to the dollar. Adoption inside the licensed population doesn't change the bill. Token-based pricing has at least four dominant variables: number of engineers, depth of engagement per engineer, model choice, and workflow shape. The first two are roughly the per-seat-equivalent multipliers. The third changes the price per token. The fourth — workflow shape — is the one corporate budgeting infrastructure is least equipped to model, because it is the one that compounds. A useful taxonomy of what workflow shape means in practice came not from any of the news coverage but from an HN commenter who organized the patterns into three escalating levels: Level Pattern Why it burns tokens Why engineers reach for it 0 Long-lived conversations without compaction or summarisation Every turn re-pays the cost of the entire accumulated context Re-establishing the agent's understanding feels more expensive than continuing the "educated" one 1 Subagent fan-out — "spawn 5 subagents to analyze your solution, each analyzing a different angle, summarize their findings" One prompt → five (or more) parallel agent runs, each consuming the seed context plus its own work Genuinely productive when used carefully; "addictive" when used at default 3 Parallel-worktree multitasking — ten Claude Code instances running across separate git worktrees, engineer cycling between them Wall-clock parallelism ≈ ten engineers' worth of token consumption per engineer Wait time on any one agent becomes work time on another Each level multiplies token consumption against headcount. Across 5,000 engineers ramping from 32% to 84% adoption in four months, the population shape becomes the bill. The Uber data is consistent with a workforce ramping into Level 1 and Level 3 patterns over four months. "Multi-step capabilities" — the framing the Briefs.co coverage attributed to the rollout's productivity gains — is exactly the workflow shape that produces the cost. Engineers discovered they could parallelize. They did. Each one's bill went from a few hundred dollars a month to a few thousand. The total moved up an order of magnitude across the organization in the same window. Goodhart's Law arrives on schedule One HN commenter named the second-order failure mode that several others described variations of: in some organizations, "rewarding high token usage as productivity without critical evaluation." The structural problem is the standard one Goodhart's Law has been describing for fifty years — when you make a metric the target, the metric stops measuring what it used to measure. Token consumption is the easiest AI-productivity metric an executive can pull off a dashboard. It is also exactly the metric that rewards Level 0 conversation-padding and Level 1 subagent fan-out whether or not those patterns are producing outcomes anyone would otherwise pay for. The Uber numbers do not, on their own, establish that this is what happened at Uber. The 70%-of-committed-code figure is an outcomes metric, not a token-spend metric, and it is not a number a Goodhart-distorted incentive structure produces. But the underlying point holds in the broader population: in a 2026 in which token usage is increasingly visible on engineering-management dashboards, the gap between tokens spent and value produced is the gap most easily mistaken for productivity. The "too valuable to afford" equilibrium The version of this question that lands hardest in the corporate-finance literature is the one Uber's CTO is now trying to answer: when a tool produces 70% of your committed code at $500–$2,000 per engineer per month, and you have ~5,000 engineers, what do you do? The math, on the cost side, does not work. The math, on the productivity side, makes turning the tools off the most expensive thing you can do. Neither answer is stable. There are three plausible re-equilibrations available to the industry, and each of them puts pressure somewhere. Flat-rate enterprise tiers — vendors offer a per-seat or per-organization price, absorbing the variance themselves — put margin pressure on the providers, who are already heavily subsidized by venture funding. Bring-your-own-API-key — the company asks engineers to expense their own API usage and reimburses up to a cap — pushes the cost into individual decisions and produces sharp visible income-bracket effects on what tools developers can use at home. Usage caps — the company sets a hard quota per engineer, expecting some workflows to fail at the cap — converts the cost problem into a productivity problem. Some mix of all three is what 2026 enterprise IT departments are about to negotiate. The Uber data is the first time the question has been put publicly by an organization at this scale. It will not be the last. What's actually new Most of the conversation around AI tooling cost in 2024–2025 was about the per-token list price — Claude is $X per million input tokens; GPT is $Y per million output tokens. That conversation is the wrong altitude. The list price has been falling, often substantially, on a six-month cadence. The thing that does not fall on the same cadence — and that the Uber data illuminates — is the adoption-shape multiplier. As tools improve, engineers run them more aggressively. The token bill is not an artifact of tool price. It is an artifact of how productive the tool is, multiplied by how many engineers are running it at near-full capability, multiplied by what workflow shapes their company tolerates. This is the bill that is unfamiliar to corporate finance, because no per-seat SaaS line item has ever behaved this way. The closest analog is bandwidth pricing in the early-cloud era: a usage-based metered cost on a resource that scaled with how successfully the underlying product was used. Companies that did not understand the cost shape under-budgeted. Companies that understood the cost shape early — Netflix's CDN strategy is the canonical example — built their cost structures around the variable they couldn't otherwise predict. The AI-tooling version of that build-around-the-variable problem is the one Uber is now navigating in public. The CTO's "back to the drawing board" phrase is the corporate-finance euphemism for we did not model this correctly the first time. The interesting question is not whether other companies will hit the same wall. They will. The interesting question is what the second-iteration AI-tooling budgeting framework looks like — and whether it can hold up under the next productivity-shape multiplier the tools enable when they get even better. What it tells you Three things stay with me from the Uber data. The first is that the pricing model is the story. The list price is a distraction. The adoption-shape multiplier — the workflow patterns engineers reach for once the tools start being useful — is the variable that turns a manageable line item into a budget-busting one in four months. The second is that the productivity numbers are real. 70% of committed code, 11% of live backend updates produced by Claude-Code-driven agents, are not the kind of numbers that appear in a Goodhart-distorted environment. Whatever else Claude Code is doing at Uber, it is producing outcomes the engineering organization will defend in front of the CTO. The third is that the too-valuable-to-afford equilibrium is unstable, and the resolution will not be graceful. The next twelve months are when the bill arrives at companies one or two orders of magnitude smaller than Uber. The CTO's confirmation is the leading indicator; the next round of vendor-pricing-model announcements is the trailing one. Token-based pricing, in its current shape, does not survive adoption curves. The Uber data is what that statement looks like with names and numbers attached.

FOMO is strong around AI, with companies adopting the technology willy-nilly. Many have encouraged employees to tokenmaxx to their hearts’ content. Now they’re starting to realize that freedom comes with a price. Just one in four companies say they have a comprehensive view of what artificial intelligence is costing them, according to an as-yet-unreleased KPMG survey reported by The Wall Street Journal. Only about half have even some visibility into the cost of their AI use. One in five have no visibility, or only see the damage once the bill arrives. “It’s a new resource that needs to be managed that didn’t exist quite that way, and we’re seeing exponential growth,” Steve Chase, KPMG’s global head of AI, told the Journal. Part of the problem is pinning down what, exactly, AI costs. The basic unit of AI use—the token—is an unusual thing to budget around. Each token is a fragment of text, code, or data processed by a model when it reads a prompt or produces an answer, but it doesn’t map neatly onto a single word. Some tokens can be cached by AI models, meaning they are not charged again, while others must be processed as new. The result is uncertainty that may not become clear until the bill lands at the end of the month. Multiply that across individual employees at a company, and it is little wonder that chief financial officers are being left with eye-watering bills. KPMG is working with companies that have blown through annual token and cloud-computing budgets in a matter of months, Chase told the Journal, while one client has seen token usage rise sixfold. Axios reported last month that one AI consultant’s client spent half a billion dollars in a single month after failing to put usage limits on employees’ Claude licenses. “People are getting these massive bills,” says Sam Ransbotham, professor of analytics at Boston College’s Carroll School of Management. “They turn on usage, and suddenly the people paying the bill are not the people using the product, and whenever you have that sort of mismatch, there’s going to be problems.” The challenge is made worse by a shift away from the “all you can eat buffet” phase of AI pricing, as vendors look to recover the enormous cost of providing access to powerful models. “Software-as-a-service for years has worked off per seat, per licensing, because companies need budget,” says Ransbotham. “They need to predict.” But AI is priced differently. The idea that AI is a game changer for business, encouraged by bosses, is also creating perverse incentives. Amazon recently shut down an internal AI usage leaderboard after employees tried to boost their scores with needless activity. Some workers reportedly assigned AI agents to pointless tasks in an attempt to climb the rankings. “The fact that you have the capacity doesn’t say that you are using it beneficially,” says Baruch Lev, Philip Bardes professor of accounting and finance at New York University Stern School of Business. Lev argues that companies need two measures: how much AI is being used, and what benefit that use produces. Without both, return on investment is mostly guesswork. “There is no resource, however beneficial, that you should use the maximum,” he says. Part of the problem is that companies still haven’t figured out how they want to account for their investments in AI. Lev says some may want to treat AI spending as an expense, akin to rent or salary. Others may see it as an investment in capacity, like software infrastructure that can be scaled across the business. Whichever answer a company’s accountants land on matters—and will likely affect the willingness of boards to keep funding AI rollouts. For companies hit with stinging bills, Lev says there is likely room to negotiate payment plans or cheaper alternatives. Some may need to pay from cash flow, borrow, or raise equity to cover the costs. But he says vendors have an incentive to be flexible. “The main vendors are willing to help,” says Lev. “They also want to show revenues on their balance sheet, so I think they’ll give them very favorable terms, at least at the beginning.” Without that flexibility, he adds, “people will not be able to pay.” To avoid future headaches, Lev says companies need to do a better job of measuring the return on their AI investments. Even without firm evidence, Boston College’s Ransbotham says it is striking that “none of them say, ‘All right, turn it off.’” Instead, he says, the reaction is more measured: “What they say is, ‘Let’s not get the Cadillac version of everything.’” That more judicious approach—using cheaper models for simpler tasks and reserving the most expensive systems for work where their capabilities actually matter—is something companies, including Coinbase, are starting to consider, according to a recent post by CEO Brian Armstrong on X. Good takeMy guess is– demand for intelligence is near infinite– but 80% of workloads will be running on 99% cheaper models within 12-18 months– 20% of workloads will still run on latest gen models where IQ maxing is important (scientific breakthroughs, higher level… https://t.co/gAFtYjorRh— Brian Armstrong (@brian_armstrong) June 8, 2026 Companies may route simple queries to cheaper models, keep multiple suppliers alive, and use the most advanced models only when the work demands them. Asking the most advanced model on the market for tomorrow’s weather is a “zillion dollar hammer for this tiny task,” says Ransbotham. Lev, for his part, pushes back against skeptics who see the big bills as evidence of an AI hype bubble coming due. He calls AI “a real revolution,” not another dotcom delusion. But even revolutions need meters.

In Q1 2026, OpenAI and Anthropic moved enterprise customers from flat-rate plans to token-based billing. The change looks administrative, but it had a direct consequence for engineering teams: the real cost of AI became visible for the first time. The market's reaction over the following two months was enough to reopen a question many considered settled: does AI actually deliver measurable ROI? What happened when the bill arrived The most documented case is Uber. The company had encouraged all employees to use agentic tools as much as possible and even ranked AI usage internally on leaderboards. The result: the entire annual budget was consumed in four months. The response was a $1,500/month cap per employee per agentic coding tool (Claude Code, Cursor, and similar). At Brex, engineers were limited to $500/week in tokens; employees outside engineering received a $5/week cap. T-Mobile temporarily capped usage at $2,000/month per user with plans to migrate to a tiered system. One unnamed company, according to Ed Zitron in "AI Is Slowing Down" (June 2026), spent $500 million on Anthropic models in a single month due to absent spend controls. These are not isolated cases. A KPMG survey reported by the Wall Street Journal in June 2026 found that only 26% of companies have a comprehensive view of their AI costs; 50% have partial visibility; and 22% only find out what they owe after the bill arrives. Steve Chase, KPMG's global head of AI, told the Journal: "It's a new resource that needs to be managed that didn't exist quite that way, and we're seeing exponential growth." The structural problem behind the spending caps The spending caps are a symptom. The root cause, as Zitron details in the same article, is that the economics of generative AI require numbers that currently seem out of reach. Anthropics has made over $330 billion in compute commitments with Google, Amazon, and Microsoft, plus another $45 billion with CoreWeave and SpaceX. To cover those commitments, it needs $174 billion in annual revenue by 2029. OpenAI is projected to burn at least $852 billion through the end of 2030 and has over $770 billion in compute commitments. The combined projected 2026 revenue for both companies sits around $60 billion, meaning they would need 496% growth by 2029. According to The Information, cited in the article, OpenAI and Anthropic account for 89% of all AI startup revenues. Outside of NVIDIA, hyperscalers, OpenAI, and Anthropic, Zitron states he cannot identify any company spending more than a few hundred million on compute. The entire sector would need to generate over $2 trillion in annual revenue by 2030 to justify the infrastructure being built. Microsoft AI CEO Mustafa Suleyman publicly stated that Anthropic's models are too expensive and that he intends to reduce Microsoft's use of them to zero — relevant given that Microsoft represents a significant share of Anthropic's customer base. What token-based billing changed for engineering teams Before the migration to token-based billing, model errors (loops, incorrect responses, reprocessing) effectively cost the end user nothing, since the cost was covered by the flat-rate plan. Starting in Q1 2026, every token consumed appears on the invoice. Zitron addresses this directly: "Think of it like this: if you're using an AI subscription with rate limits but no actual costs, any mistakes a model makes — such as getting stuck in a loop or just doing the wrong thing — can be dismissed as the troubled nature of early-stage technology, because the 'cost' was $20, $100, or $200 for the entire month. Anthropic, OpenAI and every other AI company deliberately obfuscated these costs because they knew that the second a user actually had to pay for the fuckups of an AI model they'd scream like they were being stung to death by bees." Ed Zitron, "AI Is Slowing Down" In other words, flat-rate billing acted as a buffer that made model inefficiency invisible. With token-based billing, every unnecessary agent iteration appears on the invoice. This places new requirements on engineering teams adopting agentic tools: Cost observability: real-time dashboards showing token consumption per team, per tool, per task type Preventive limits: spending caps configured before scaling adoption, not after receiving the first surprise invoice Output metrics: connecting token spend to measurable results (pull requests merged, tickets closed, features shipped to production) Agentic workflow review: identifying unnecessary loops or reprocessing steps that consume tokens without producing different output What to evaluate before scaling adoption The shift to token-based billing does not make AI tools less useful. What it does is make the cost honest. For teams that do not yet have this instrumentation, the risk is the same as Uber's: rapid adoption growth, budget exhausted in months, and a forced decision to cut access rather than optimize usage. We should evaluate, workflow by workflow, which use cases have verifiable ROI and which are in experimentation. The difference between the two is not a judgment call, but an operational one: experimentation needs a separate budget and explicit exit criteria. How is your team handling cost visibility for agentic tools? Fonte: AI Is Slowing Down — Ed Zitron