The Meter Drop

On April 27, GitHub announced that all Copilot plans will transition from flat-rate subscriptions to usage-based billing on June 1, 2026. The blog post's second sentence: "Plan prices aren't changing." The blog post's pricing tables tell a different story.

Copilot Pro still costs $10 a month. Under the new system, that $10 buys $10 in "AI Credits." One AI Credit equals one cent. A single Opus 4.6 request that previously counted as one premium request now costs tokens at $5 per million input, $25 per million output. A heavy coding session with Opus can burn through $10 in credits in an afternoon.

One HN commenter reported their own usage. Over two months of Copilot Pro, they estimated they consumed roughly $1,900 in tokens at on-demand pricing, while GitHub charged them $22 in premium requests. "They are undercharging by a factor 50 to 90," the user wrote. "This is just the start of the rug-pull."¹ The exact figure depends on assumptions about token rates and caching, but the scale of the subsidy is not in dispute.

The multiplier table

For annual subscribers who keep their current plans, GitHub published a model multiplier table that reveals the scale of the subsidy they've been absorbing. Here are the numbers.

Claude Sonnet 4.6 goes from 1x to 9x. Claude Opus 4.6 goes from 3x to 27x. GPT-5.4 mini goes from 0.33x to 6x, an 18-fold increase. GitHub was eating the difference between what it charged and what inference actually cost, and the table is an admission that the gap was enormous.²

Every model that developers actually want to use just got dramatically more expensive. The models that stay cheap are the ones nobody was choosing.

The credit illusion

GitHub's framing is careful. "Your plan pricing is unchanged." This is technically true in the same way that a gym membership is unchanged when they remove every machine except the treadmill.

Under the old system, Copilot Pro included 300 premium requests per month, with unlimited completions. Under the new system, code completions remain free, but everything else runs on the meter. Chat, agents, code review, multi-file edits. The features that defined the upgrade from autocomplete to AI pair programmer now have a visible, per-token price.

As one commenter put it: "Isn't this like saying 'The Porsche you rented at $200/mo is now a Honda. But the price hasn't changed!'"¹

The pattern

This is the fifth chapter in a sequence we've been tracking since March.

Anthropic rationed Claude quietly, then confessed, then started charging for the overflow. We wrote about it in The Rationing, Pay-Per-Overage, and The Credit. GitHub did the same thing on a longer timeline: introduced rate limits in April, removed models from cheaper tiers, paused signups, offered refunds. We wrote about that in Copilot's Ten-Day Cliff and Copilot Credits.

Now GitHub has changed the billing model itself. The flat-rate era of AI coding tools is ending because it was never economically viable. Every vendor subsidized inference to acquire users, and every vendor has now found a way to stop.

Windsurf made a similar move in March, replacing credits with daily and weekly quotas.³ As Simon Willison noted in the HN thread: "With hindsight, per-request pricing makes no sense."¹

The enterprise bet

The one group that benefits from this change is enterprise finance teams. GitHub is introducing pooled credit usage across organizations, granular budget controls at the enterprise, cost center, and user level, and promotional credits ($30/month for Business, $70/month for Enterprise) through August.⁴

This is the play. Individual developers who were burning through hundreds of dollars in subsidized inference on a $10 plan will leave. Enterprise customers who need per-seat budget controls and compliance reporting will stay, because the alternative is giving every developer a raw API key and hoping for the best.

GitHub is betting that enterprise stickiness absorbs the individual churn. Based on the multiplier table, they need it to. A heavy Opus user on a $10 plan was costing them somewhere between $500 and $1,900 per month in actual inference, a customer acquisition cost that never converts.

What happens June 1

Monthly subscribers auto-migrate. Annual subscribers keep their current pricing until renewal, though model multipliers still increase on June 1. Fallback experiences to lower-cost models are discontinued entirely. Code review will consume both AI Credits and GitHub Actions minutes.⁴

The practical effect: developers who used Copilot as an unlimited AI partner will need to watch a meter. Developers who used it for completions only won't notice. The gap between those two groups is the entire value proposition of the last two years of Copilot marketing.

Several commenters in the HN thread (668 points, 488 comments and climbing) announced they were canceling immediately. "Cancelling. Going with Codex $100, Kimi annual plan, DeepSeek API, and a local LLM once I get a Mac Studio," wrote one. Another: "If there's no discount on credits over other providers, I'm going to switch to a PAYG provider."¹

Usage-based billing is probably the right model for AI coding tools. The transition from "unlimited" to "metered" cannot happen without developers feeling like the product they were sold no longer exists. Based on 488 comments, most of them agree.

Share on Bluesky · Share via Email