The Rationing: Copilot Credits

What Changed

The prices look identical. Pro is still $10 a month. Pro+ is still $39. Business is $19 per seat. Enterprise is $39. What changed is what those prices buy. Under the old system, you paid a flat rate and got a bucket of "premium requests." Under the new system, each plan comes with a matching dollar amount of AI Credits. Pro gets $10 in credits. Pro+ gets $39. When the credits are gone, Copilot stops working until the next billing cycle. Credits are consumed based on token usage: input tokens, output tokens, and cached tokens, priced at published API rates per model. This is token billing with a friendlier name. As one Hacker News commenter put it: "$10 of monthly credits is basically just a prepaid wallet with a reset timer." Code completions and Next Edit suggestions remain free. Chat and agent features, the parts developers actually value in 2026, consume credits. The old system had "fallback experiences," degraded modes that kept working when you hit limits. Those are gone. When credits run out, the tools stop.

The Multiplier Shock

For annual plan subscribers, the model cost multipliers changing on June 1 tell the real story: Opus 4.6: from 3x to 27x. A nine-fold increase. Opus 4.7: from 3x to 27x. GPT 5.4: from 1x to 6x. GPT 5.4 mini: from 0.33x to 6x. An eighteen-fold increase. Sonnet 4.6: from 1x to 9x. One Hacker News commenter calculated what this means in practice: "On my $10 plan, running either 100 Opus or 300 Sonnet agent runs would cost hundreds of dollars." Another: "27x for Opus is genuinely shocking. At that point you're not paying for convenience anymore, you're just paying a GitHub tax." Annual plans are being retired. Existing annual subscribers stay on the old pricing until expiration, then auto-downgrade to Free. GitHub is not offering annual renewals.

One Week After the Cliff

This announcement came exactly one week after GitHub's emergency rationing on April 20. That day, GitHub pulled Opus from the Pro tier, paused new signups for all paid plans, tightened usage limits to two ceilings (session and weekly), and opened a refund window through May 20. The blog post blamed "agentic workflows" for changing compute demands: "long-running, parallelized sessions now regularly consuming far more resources than the original plan structure was built to support. It's now common for a handful of requests to incur costs that exceed the plan price." The April 20 rationing was a bridge measure. The credit system was already in development. The emergency tightening bought GitHub five weeks to finalize the billing overhaul. The sequence was: subsidize, discover the subsidy is unsustainable, panic-ration, then formalize the rationing into a billing model.

The Pattern

GitHub is the fourth major AI coding platform to abandon flat-rate pricing for usage-based billing in under a year. Cursor replaced fixed request allotments with credit pools in June 2025. One user reported $350 in overage charges in a single week. Cursor issued a public apology on July 4 and offered refunds. Anthropic spent seven weeks in March and April 2026 dealing with three separate bugs that drained Claude Code users' quotas. One user documented sessions generating 652,000 tokens without any user prompts. Anthropic's postmortem admitted their internal teams were not required to use the customer-facing builds. Windsurf moved to quota-based pricing in March 2026. GitHub's announcement makes it four for four. Every platform that launched with generous flat rates to capture market share has discovered that agentic usage burns 10 to 100 times more tokens than the chat workflows the pricing was designed for. Simon Willison, commenting on the Hacker News thread: "With hindsight, per-request pricing makes no sense at all if an agent can burn a widely varying amount of tokens satisfying that request. These pricing plans were designed before coding agents changed the dynamics of token usage."

The Same Day

On the same day GitHub announced Copilot credits, Microsoft published a blog post restructuring its financial relationship with OpenAI. Microsoft will no longer pay a revenue share to OpenAI for reselling OpenAI products. In exchange, Microsoft drops its exclusive right to sell OpenAI's models. GitHub is owned by Microsoft. OpenAI's models power much of Copilot. The revenue restructuring means Microsoft keeps more margin on every Copilot token consumed. The usage-based billing switch makes that margin directly proportional to consumption rather than absorbed by flat-rate heavy users. The timing may be coincidental. The financial logic is not.

The GitHub Tax

Multiple Hacker News commenters pointed out that Copilot's per-token pricing, once the credit abstraction is removed, is significantly more expensive than going directly to the API providers. For GPT 5.5, Copilot charges 7.5 times what GPT 5.4 costs at the API level, while OpenAI's own Codex product charges only 2 times. The difference is the GitHub markup: authentication, IDE integration, context management, and the convenience of not managing your own API keys. That convenience premium was invisible under flat-rate pricing. Under usage-based billing, every developer can calculate exactly what they are paying above the API rate. Several commenters concluded they would switch to OpenRouter, direct API access, or Claude Code. Others noted they are locked in: "Because Copilot is the only thing allowed at our corp." Enterprise purchasing decisions do not respond to individual developer preferences. That is the moat.

The Inversion

Code completions, the original Copilot product from 2021, remain free. Chat and agent features, what developers actually want in 2026, cost credits. This is the classic platform inversion. The product that built the user base becomes the loss leader. The product that users depend on becomes the revenue engine. Multiple Hacker News commenters said they would keep Copilot solely for free autocomplete and do their real work in Claude Code or other tools. The developer who pays $10 a month for Copilot Pro gets free code completions, $10 in agent credits that run out quickly, and the knowledge that the same models are available at lower per-token rates through direct API access. The value proposition has shifted from "pay for a service" to "pay for the convenience of not switching."

What Comes Next

The Hacker News thread surfaced a prediction that recurred across dozens of comments: local and open-source models will capture the developers that usage-based pricing pushes out. Qwen, DeepSeek, and Llama were all mentioned as crossing the usability threshold for coding in early 2026. One commenter estimated that local models save $360 to $1,800 over three years compared to subscriptions. We wrote about this two days ago in The Weekend Developer. The tools exist. The models are good enough. The only question was when the pricing would push developers to actually switch. June 1 is the answer.