Two separate announcements in two days, from two different companies, for the same reason.
On April 20, GitHub announced changes to Copilot individual plans: new sign-ups for Pro, Pro+, and Student plans paused, usage limits tightened, and Opus model access removed from Pro-tier subscribers. The stated rationale was direct — “long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support.” A single agentic session, they noted, could incur costs exceeding the monthly plan price.
On April 22, Anthropic’s pricing pages and support documentation were updated to show Claude Code as available only on the Max plan ($100/month), not Pro ($20/month). Anthropic characterized it as “a small test on ~2% of new prosumer signups.” After several hours of public backlash the change was reversed, but the support documentation update — “Using Claude Code with your Max plan” — was not the kind of thing you write accidentally for a two-percent test.
The common thread is that both companies built their subscription economics around copilot-style interactions: fast completions, bounded context windows, requests that finish in seconds. A developer asking Copilot to autocomplete a function is cheap. A developer asking a coding agent to implement a feature, debug the test suite, and open a PR is not. The latter can run for thirty minutes, consume tens of thousands of tokens across dozens of tool calls, and parallelize across sub-agents — potentially using more compute in a single session than a typical user would consume in a month under the old interaction model.
The math stops working. At $20/month, you can sustain a lot of autocomplete. At $20/month with agents, you’re subsidizing arbitrarily expensive sessions for the heaviest users while lighter users fund the difference. GitHub’s solution is to pause growth and tighten limits while they figure out sustainable pricing. Anthropic’s briefly-attempted solution was to route the expensive tool (Claude Code) to the expensive plan.
The Anthropic reversal is interesting in its own right. They backed down quickly, but the underlying pressure doesn’t go away. Claude Code is one of the more compute-intensive ways to use Claude — a typical agent loop involves many reasoning steps, long context reads, and repeated tool calls. Running that against the Opus 4.7 or Sonnet 4.6 backend on a $20/month plan is not a sustainable business model at scale. The reversal was probably the right short-term call (losing paying Pro users is bad) but it’s a delay, not a resolution.
The likely endpoint is some form of compute-time or token-based billing layered on top of the base subscription. GitHub has reportedly been evaluating token-based billing for a while; the current plan changes are probably a bridge while that system is built. Anthropic already distinguishes Max from Pro partly on compute capacity; extending that logic to specific tools is a natural next step.
What’s actually interesting here is timing. Both of these companies are bumping into the economic limits of flat-rate pricing precisely when agentic use is becoming genuinely valuable — not as a novelty but as something developers use daily. The tools are useful enough that heavy users will pay more; the question is how the transition happens without driving those users to competitors (or to self-hosted models, which have no recurring cost after setup).
The GitHub changes come with a refund policy through May 20 for subscribers who decide the new limits don’t work for them. That’s a responsible way to handle a hard change. But the underlying story is that the pricing model that got AI coding assistants to mass adoption is already obsolete, and both GitHub and Anthropic are improvising toward something that doesn’t exist yet.