The main demand signals for artificial intelligence look explosive on paper, but they may be significantly exaggerated. By pricing its tools to match that reality, Anthropic could be the best-positioned AI company with an adjustment.
Tokens are the basic unit of AI usage, the words and characters that make up both the queries that users submit and the output models that are generated.
Chatting with the AI consumes hundreds of tokens per paragraph. With Agentic AI, models write code, browse the web, execute multi-step workflows, and execute thousands of additional workflows per session.
Using Anthropic’s latest model rates, an input (prompt) of 1 million tokens costs $5, and an output (model response) of 1 million tokens costs $25.
AI companies cite the boom in token consumption to justify the hundreds of billions of dollars spent on the infrastructure that services them.
However, token consumption is becoming a distorted indicator.
Meta and Shopify It says it has created an internal leaderboard to track the number of tokens used by employees. Nvidia CEO Jensen Huang said he would be “very alarmed” if an engineer making $500,000 a year wasn’t using at least $250,000 worth of computing. It means measuring how much money engineers spend on AI, not what they produce with it.
As companies begin to measure AI adoption, employees will optimize for metrics rather than outcomes.
“If you just want to burn a lot of money, there are easy ways to do it,” said Ali Ghodsi, CEO of Databricks, which processes AI workloads for thousands of companies. “You resend the query to 10 places, and you create a loop that just repeats it over and over again. That costs a lot of money and doesn’t yield anything.”
Jen Stave, executive director of the Harvard Business School AI Institute, hears the same thing from corporate leaders.
“I’ve talked to about a dozen CTOs and CIOs, and they all say, ‘Actually, we’re having a really hard time finding an ROI framework for this,'” she said.
Anthropic plans for the possibility that demand forecasts are wrong.
CEO Dario Amodei described a situation he called a “cone of uncertainty.” Data centers take one to two years to build, so companies are currently shelling out billions of dollars for unmet demand. If you don’t have enough capacity, you will lose customers if you buy too little. If you buy too much and your returns don’t arrive on time, the calculations will stop working.
“If it’s a few years off, it could be ruinous,” Amodei said on Dwarkesh Patel’s podcast in February. “I get the impression that some other companies don’t write down spreadsheets. They just do it because it sounds cool.”
Anthropic’s response is to move from flat-rate enterprise pricing to per-token charging, so the revenue Anthropic collects reflects actual usage. OpenAI has also made AI cheaper and more accessible at scale, while also discontinuing some third-party tools that were big consumers of tokens.
In the early days of AI adoption, flat fees were the norm, with generous or unlimited AI access available for a fixed monthly fee. This model worked when people were chatting with the AI. However, using agents changed the cost from thousands of tokens per session to millions, disrupting the economic structure.
Anthropic’s most generous consumer offering, the $200/month Max plan, served as a case study.
Developers routed that subscription through a third-party agent tool like OpenClaw and had AI agents running around the clock based on plans designed for conversations. Based on Anthropic’s published pricing for the latest model, heavy users of Claude Code Max can pay as little as $200 a month for usage that would cost up to $5,000 without a subscription.
On April 4th, Anthropic retired those tools. Boris Cherny, head of Claude Code, wrote in X that subscriptions are “not built for the usage patterns of these third-party tools.”
The same readjustment is happening in businesses.
Previous Anthropic contracts included standard seats and premium seats (a flat monthly fee with usage built-in). According to the company’s support page, these are currently listed as “legacy seat types no longer available for new Enterprise contracts.” The new Enterprise plan is priced per seat and has an API rate added to token consumption.
Anthropic was the first to make the move, but pressure is building across the industry.
OpenAI’s Nick Turley, head of ChatGPT, acknowledged on the BG2 podcast, “In this day and age, having an unlimited plan can be like having an unlimited power plan. It just doesn’t make sense.”
Once every token has a price, businesses and consumers who have budgeted for subscription AI will start asking what they are actually getting in return.
Ramp CEO Eric Greiman, who recently launched a token tracking tool, sees the move from a financial perspective.
AI spending across Ramp’s customer base has increased 13x in the past year, and no one knows how to budget it. He pointed to Anthropic’s approach as a smarter long-term strategy and raised questions that should concern OpenAI investors. So if your business model relies on extracting maximum token spend, is there any incentive to help your customers use AI more efficiently?
Salesforce is making a similar bet, rolling out a new metric called “Agent Work Units” that tracks work completed by AI rather than tokens it burns.
Both Anthropic and OpenAI are expected to aim for IPOs this year. The first question public market investors will then try to answer will be about demand.
By moving to per-token charging, Anthropic will be able to get clearer data on what customers actually value. OpenAI’s numbers will be large, but it will be difficult to prove how much of them are real.
If even a significant portion of today’s AI demand inflates, the companies that price with reality in mind will be the ones that survive when the correction arrives.
