The Time Bomb Went Off: AI's All-You-Can-Eat Era Just Ended in Real Time

Subscribe

All Articles

AI & Technology

The Time Bomb Went Off: AI's All-You-Can-Eat Era Just Ended in Real Time

May 17, 2026

Six days ago, we published a piece arguing that every AI subscription is a ticking time bomb for enterprise. We did not expect the fuse to be quite this short.

Share:

Copied!

Image Credit: State of Brand

Six days ago, we published a piece arguing that every AI subscription is a ticking time bomb for enterprise. The argument: AI labs are selling compute at a fraction of its cost, companies have built load-bearing workflows on those subsidized prices, and when the correction comes, it will be brutal.<br>We did not expect the fuse to be quite this short.<br>On May 14, three days after that article went live, both Anthropic and OpenAI made moves that took the repricing from theoretical to actual. It is happening now, publicly, with competitive aggression that tells you neither company thinks the current model survives the year.<br>What Happened on May 14<br>Anthropic announced that effective June 15, it is splitting its Claude subscription into two separate usage pools. Interactive usage, the kind where a human sits at a keyboard and has a conversation, stays covered under existing subscription limits. But automated and programmatic usage, the kind generated by agentic tools like OpenClaw and third-party harnesses that route tasks through Claude around the clock, is being pulled out of the base subscription entirely and placed behind a separate credit meter with fixed monthly API credits.<br>That is the structural change the original article warned about. The all-you-can-eat model did not survive contact with agentic workloads. Anthropic's head of Claude Code, Boris Cherny, had already said in April that subscriptions "weren't built for the usage patterns of these third-party tools." The June 15 change makes that position official policy. Axios put it well: the industry is rediscovering a lesson from earlier eras of computing. Humans have built-in limits to how much data they can consume. A human might send dozens or hundreds of prompts a day. An autonomous coding agent can generate thousands of requests, run tests continuously, browse the web, and recursively call models. The subscription was priced for the human. The agent showed up anyway.<br>The reaction from users was immediate and hostile. Critical replies flooded the announcement, with developers calling the changes "gaslighting" and publicly declaring they were switching to OpenAI's Codex. Which is exactly what OpenAI was counting on.<br>Within hours of Anthropic's announcement, Sam Altman posted that OpenAI would give enterprise users two free months of Codex access if they switched within 30 days, complete with a one-click migration tool that transfers Claude Code prompts, skills, and MCP configurations. Anthropic countered the same day by boosting Claude Code's weekly usage limits by 50% for Pro, Max, Team, and Enterprise users, valid until July 13, a window that overlaps almost perfectly with OpenAI's free trial period.<br>Call it what it is: a pricing war fought in the open, with each company trying to lock in developers before the IPO window opens in the second half of the year.<br>And they are not the only ones moving. GitHub announced that all Copilot plans will transition to usage-based billing on June 1, 2026, replacing flat-rate premium requests with token-based AI Credits. Their own blog post was unusually direct about the reason: "Copilot is not the same product it was a year ago. It now powers far more complex, agentic workflows that consume far more compute." Annual plans will not auto-renew. Monthly subscribers will be migrated automatically. The flat-fee era at GitHub is over in two weeks.<br>The Stealth Hike Nobody Noticed<br>The subscription restructuring grabbed headlines. But a quieter change had already been extracting more money from developers for weeks.<br>When Anthropic released Claude Opus 4.7 in April, it kept the per-token rates identical to Opus 4.6: $5 per million input tokens, $25 per million output tokens. Same numbers on the same pricing page. But Opus 4.7 shipped with a new tokenizer. For the unfamiliar: the tokenizer is the component that sits between your text and the model and decides how many tokens your words are worth. The new one is more aggressive. The same input, the same sentence, the same document, now generates up to 35% more tokens than it did on the previous model. For developers who migrated to the new flagship without benchmarking, bills climbed by as much as 27% with no visible change to the pricing page.<br>That kind of repricing does not show up in an announcement blog post. It shows up on the invoice three months later.<br>The Enterprise Contracts Already Changed<br>The consumer subscription changes are the part you can see. Underneath, the enterprise contracts have been shifting since late 2025.<br>The Register reported in April that...