What Happened (the pricing page as a mood ring)

The Register spotted a small but loud change: Anthropic’s Pro plan briefly stopped showing Claude Code as included, while other parts of the site still suggested it was. Anthropic’s head of growth later said it was a test affecting around 2 percent of new “prosumer” signups, and existing Pro/Max subscribers weren’t impacted.

This is the modern way to run a pricing experiment: change a line of HTML, then watch Reddit experience a collective out-of-body event.

Why This Matters (subscription pricing can’t survive agentic usage)

Coding agents are not “chat.” They are workloads. They run longer, they take more steps, they burn more tokens, and they spike demand at exactly the moments everyone else spikes demand (weekday business hours, product launches, demo season, and, naturally, whenever a new model gets hype-trained by Twitter).

Anthropic’s explanation, as quoted by The Register, is basically: we designed plans for heavy chat, and then the product became long-running agents (Claude Code, async agents, “Cowork”). Engagement per subscriber went up. The plans did not evolve fast enough. So now they’re testing options.

This is not an Anthropic-only problem. It’s an industry problem with a simple equation: if the book value of tokens consumed is materially higher than the subscription fee, you either raise prices, cap usage, or change what’s bundled. You can call it “plan optimization.” I call it “math.”

The Capacity Backdrop (Amazon just bought the next decade)

Right on cue, the infrastructure layer shows up with a checkbook. Anthropic announced a compute deal tied to Amazon investment, promising “meaningful compute in the next three months” and nearly 1GW before the end of 2026, plus access to multiple generations of Trainium chips. Ars Technica frames it as circular financing: Amazon funds Anthropic, Anthropic spends on AWS and Amazon silicon.

That deal is not just growth fuel. It’s a symptom: consumer and developer adoption is now fast enough to break reliability. When reliability breaks, pricing experiments start appearing like mushrooms after rain.

So What Is This, Really?

This is the moment where frontier labs discover that “developer love” is a cost center unless you can meter it. The developer tooling layer is strategic substrate, but it’s also the easiest place to accidentally subsidize the most intense users. Claude Code users are not casually asking for poems. They’re running the model like it’s a colleague with infinite overtime.

And enterprises hate uncertainty. The Register is blunt about the comms incoherence: if the public pricing page changes, the internet notices, even if you intended to only test new signups. If you want trust, you need predictable contracts, explicit notice, and a plan hierarchy that makes sense.