OpenAI and Anthropic have done the same thing to their biggest customers. As of April 2026, both bill enterprise Claude Code and Codex usage at standard API token rates instead of the discounted flat fees that fueled the first wave of coding agents. Two more expensive frontier models landed the same month.
Same move, two timelines
OpenAI was the explicit one. Its Codex rate card says that on April 2 the company swapped per-message billing for per-token pricing across new and existing Plus, Pro and Business plans, plus new Enterprise accounts. On April 23 it extended that to every remaining Enterprise contract, including the education, health and government tiers.
Anthropic was quieter. It began moving enterprise accounts to per-token billing back in November 2025, according to a spokesperson, but plenty of customers only noticed when their renewals came up this spring. The seat now buys platform access and little else: $20 a month for a Claude Code seat, $10 for a Claude.ai one, with every token billed on top. The old 10 to 15 percent volume discounts are gone.
The sticker price went up too. Sort of.
April also brought pricier models. GPT-5.5 arrived on April 23 at double the per-token rate of GPT-5.4, climbing to $5 per million input tokens and $30 per million output. Anthropic's Opus 4.7, out April 16, kept its $5 and $25 rate card untouched, which sounds generous until you hit the new tokenizer.
That tokenizer can turn the same prompt into as much as 35 percent more billable tokens. Willison pegs the real increase at roughly 1.4 times Opus 4.6. OpenAI plays the same trick in reverse, arguing GPT-5.5 burns fewer output tokens, so the effective bill rises closer to 20 percent than 100. Both figures come from the companies selling the tokens. Treat them as marketing until someone independent runs the workloads.
Why now?
Simon Willison, whose product-market fit post kicked off this round of analysis, has a blunt read: the labs finally found something people will pay real money for. He ran the ccusage tool against his own heavy month and found he had burned about $2,180 of API tokens on $200 of subscriptions. Companies running agents across thousands of engineers do not get that arbitrage.
The revenue math backs the theory. Anthropic told investors it expects $10.9 billion in second-quarter revenue, more than double the $4.8 billion it booked in Q1, and a first operating profit of around $559 million. CNBC and the Wall Street Journal both reported the numbers on May 20.
Read the fine print on that profit
The profit claim deserves a squint. Anthropic's first quarter in the black happens to be the same quarter it locked in cheaper compute. A SpaceX compute deal disclosed in the rocket maker's IPO filing has Anthropic paying about $1.25 billion a month for Colossus capacity, with a reduced rate for May and June while the deal ramps. Two of the three months in that profitable quarter run on a builder's discount. Writer Ed Zitron read the timing as an accounting quirk dressed up as a milestone, and the Journal itself flagged that profitability may not hold once full compute costs arrive.
The "AI is failing" stories are thinner than the headlines
Some of the loudest coverage this spring framed all this as companies drowning in surprise bills. Uber maxed out its full-year 2026 AI budget within a few months, mostly on Claude Code spreading across roughly 5,000 engineers, and CTO Praveen Neppalli Naga told The Information that "the budget I thought I would need is blown away already." Which is what tends to happen when a tool gets good enough that 5,000 engineers actually reach for it.
The more quoted worry came from Uber COO Andrew Macdonald, who in podcast comments struggled to connect Claude Code usage to shipped features. "That link is not there yet," he said. Read it cynically and that is an admission AI is wasted money. Read it Willison's way and it is a buyer wincing at the price before signing the renewal anyway. Microsoft, separately, has been canceling licenses for Claude Code in favor of its own Copilot CLI, a call The Verge reports is partly financial, timed to the June 30 close of its fiscal year.
The real numbers arrive with the paperwork. Anthropic is weighing an IPO as soon as October, and OpenAI has a confidential S-1 of its own in progress. Audited filings will show whether April's repricing produced durable margins or just a well-timed quarter. Every enterprise renewal between now and then is its own data point.
