Four Companies, One Pricing Collapse: What's Really Going On With AI Coding Agent Costs

OpenAI shifted Codex from per-message to token-based credit pricing on April 2, 2026. Anthropic doubled Claude Code's rate limits in May, then split off automation and API use into separate fixed credits two months later. Cursor rewrote its pricing model back in 2025 after heavy users saw effective cost increases of 20 to 70 times their subscription price. GitHub Copilot is moving to usage-based billing this June, ending a subsidy that let heavy users consume three to eight times their subscription's token value while GitHub absorbed the difference.

Four competing companies, four different pricing philosophies, and all four landed on the identical sequence within about a year of each other: launch a flat, generous-feeling subscription, discover that a small cohort of power users is consuming far more in real inference cost than they're paying, then walk it back through throttling, credit splits, or usage-based billing.

The convergence is the story, not any single company's decision

Anthropic's case shows how sharp the mismatch can get. A class action filed in the Northern District of California alleges that one five-hour Claude Code session under a $200-a-month "Max 20x" plan consumed 15% of the plan's weekly quota, far more than the advertised multiplier implied was possible. Cursor's 2025 pricing change was rooted in the same math: a shift from 500 fast requests a month to "$20 of usage at API rates" meant some heavy agentic users saw weekly overages of roughly $350, prompting CEO Michael Truell to publicly apologize and issue refunds.

The underlying numbers explain why this kept happening across companies that otherwise compete hard against each other. AI-native product gross margins sit around 52% in 2026, up from 41% in 2024, but still 23 to 33 points behind the 75 to 85% margins mature SaaS companies run. Inference cost as a share of spend is rising, not falling, as these products mature, the opposite of the cost curve software companies are used to, and the most engaged users of an agentic coding tool are mechanically the most expensive ones to serve, not the most profitable.

That inversion is what makes agentic AI products structurally different from the SaaS pricing model everyone involved, including the AI companies themselves, initially tried to apply to them. A flat $20-a-month subscription works when marginal cost per user is close to zero, which is true for most software. It does not work when your best, most active users are running the equivalent of a small server farm's worth of inference every session.

None of these four companies figured this out in advance. They found it the same way: launch flat pricing to win adoption, absorb the loss on power users as a customer-acquisition cost, then unwind it once the math became unsustainable, backlash and litigation included. That sequence, repeating identically across competitors, is the actual signal here. It suggests the "$20-a-month coding agent" was never a viable standalone business, just a venture-subsidized price point, and its industry-wide unwinding in 2026 is a preview of what's coming for every other category of agentic AI product built on the same flat-rate assumption.

Sources: OpenAI Help Center · Fortune · Yahoo Finance · TechCrunch · Tech Times · Fortune