The flat-rate coding plan is dying. The receipts say so.

Copilot moved to metered AI Credits on June 1. Google stopped serving consumer Code Assist on June 18 and pointed individuals at Antigravity. Qwen closed its cheapest tier in March. Every 2026 billing change has moved toward a meter — and none has moved back.

Since March, every billing change among the major coding-plan vendors has moved in one direction: away from a flat monthly price, toward a meter, a pool, or a pricier tier. Alibaba closed Qwen’s Lite plan to new subscribers on March 20. GitHub switched Copilot to usage-based AI Credits on June 1. Google stopped serving Gemini Code Assist’s individual tiers on June 18. Cursor split every Teams seat into two metered pools, effective July 1 for renewing customers. No vendor converted a meter back into a flat plan.

That leaves Anthropic and OpenAI as the only majors still selling a true flat window — a fixed monthly price for a message quota that resets every five hours, with no dollar or credit meter underneath (Claude pricing, Codex pricing). For a heavy user, at the index’s default assumptions, those plans compute to $0.26–$0.51 per million tokens against $3–$6 for the metered pools. The rest of this piece is the receipts and the math.

Four receipts, one direction

Between March 20 and July 1, 2026, four vendors changed how coding-plan usage is bought, and every change tightened it. The dates below come from vendor changelogs and pricing pages, not from coverage of them.

Date	Vendor	What changed
Mar 20 / Apr 13	Alibaba (Qwen)	[Lite plan closed](https://www.alibabacloud.com/help/en/model-studio/coding-plan) to new subscriptions; renewals and upgrades ended April 13. Only the $50/mo Pro tier remains on sale.
Jun 1	GitHub Copilot	[Usage-based billing live](https://github.blog/changelog/2026-06-01-updates-to-github-copilot-billing-and-plans/). Plans are now denominated in AI Credits: [Pro is 1,500 credits ($15 value) for $10/mo; Max is 20,000 credits ($200 value) for $100/mo](https://docs.github.com/en/copilot/get-started/plans).
Jun 18	Google	Gemini Code Assist IDE extensions and Gemini CLI [stopped serving individual, free, AI Pro, and AI Ultra requests](https://developers.google.com/gemini-code-assist/docs/deprecations/code-assist-individuals); the official path forward is the [Antigravity CLI](https://developers.googleblog.com/en/an-important-update-transitioning-gemini-cli-to-antigravity-cli/).
Jun (Jul 1 renewals)	Cursor Teams	Every seat [split into two pools](https://cursor.com/blog/teams-pricing-june-2026) — “Composer and Auto” vs “Third-Party API.” Standard rose to $32/mo annual ($40 monthly); Premium to $96/mo annual ($120 monthly).

The Copilot conversion is the cleanest of the four, and it landed hard. The Register reported exit threats within a day of the June 1 switch. By June 19 the metering was fully auditable: GitHub shipped per-user ai_credits_used in the Copilot usage metrics API, so an admin can now read each developer’s burn rate off an endpoint. A seat price became a meter, and the meter got a dashboard, inside three weeks.

Two mechanisms, not one

Only one of these four events is a literal flat-plan-to-meter conversion, and it matters for what you should predict next. Our own changelog classification tags Copilot’s June 1 switch as the sole meter conversion of the window and Google’s as an access change. Qwen sits outside that ledger, but its move is the third mechanism — an entry-tier removal.

Meter conversions put a price on each unit of an existing plan. Copilot is the clean case. Cursor’s Teams change is adjacent — its pools were already dollar-denominated, and the June repricing split and raised them.
Access closures delete the cheap way in without metering anyone. Google didn’t meter Code Assist individuals — it stopped serving them. Alibaba didn’t reprice Lite — it closed the door and left only the $50 Pro tier on sale.

The Qwen case is worth being precise about, because it cuts against a lazy version of this thesis. Qwen Pro computes to roughly $0.10 per million tokens at the index’s defaults — cheaper per token than any Anthropic or OpenAI tier. Qwen didn’t get expensive; it got a $50 minimum. What disappeared in March was the cheap flat entry point, and that is the pattern all four events share: the low-commitment, fixed-price way to buy coding tokens keeps getting removed.

What the meter costs, per token

At the index’s default assumptions — 25,000 tokens per interaction, 40 active hours a week, quota fully used — the metered plans cost six to twelve times more per million tokens than the flat windows. These are computed values, monthly price divided by estimated included tokens, from plan data verified through July 1.

Plan	Billing shape	Price	Effective $/M tokens
Cursor Pro	dollar pool	$20/mo	≈$6.02
Copilot Pro	AI Credits	$10/mo	≈$4.01
Copilot Max	AI Credits	$100/mo	≈$3.01
Claude Code Pro	flat 5h window	$20/mo	≈$0.51
Claude Max 20x	flat 5h window	$200/mo	≈$0.26

The gap does not close as you spend more. Copilot Pro to Claude Pro is a ~7.8× spread at $10–$20/mo; Copilot Max to Claude Max 20x is still ~11.6× at $100–$200/mo. And the assumption doing the work here is stated on the page: this is the rate for a heavy user who saturates the quota. A developer who sends a handful of messages a week inverts the math — a meter scales down with light usage, a flat plan’s fixed price doesn’t. Move the sliders and watch which side of the line your usage lands on.

Meters also carry a failure mode flat windows can’t have: unbounded downside. A Cursor forum thread documents a single session costing $600+ on the dollar-metered pool. On a flat window the equivalent incident costs you a five-hour wait.

Flat is cheaper. It is not calmer.

The honest version of this argument does not claim flat-rate users are happier — the sentiment data says they aren’t, much. In our current provider panel, Copilot scores 1/5 on quota satisfaction and 2/5 on value after the metering switch, with “usage-based pricing shock” the top complaint. But Claude Code scores 2/5 on both of those axes, and Codex — also a flat plan — beats Claude Code at 3/5 on each. The price-per-token gap is 6–12×; the experience gap is much smaller.

Anthropic’s window had its own trust incident in the same three weeks. A counter bug — acknowledged by Anthropic on June 19 and tracked in claude-code#69904 — showed an incorrect weekly limit and blocked messages for roughly 3% of Max and Pro users before Anthropic reset the affected accounts’ 5-hour and weekly counters. A flat window is only as good as the counter behind it. The structural difference is that when Anthropic’s counter misbehaves you lose access; when a meter misbehaves you lose money.

How long does the window stay open?

As long as it exists, the flat window is the anomaly to exploit — for a saturating heavy user it is the mispriced side of the market by 6–12×, per the table above. Anthropic still sells Pro at $20 and Max at $100/$200; OpenAI still sells Codex on ChatGPT tiers at $20/$100/$200 with 5-hour message quotas and no credit meter. Nothing in either vendor’s changelog suggests a conversion this quarter.

But the 2026 record is four tightening events, zero loosening ones, and the two vendors with the deepest per-token subsidy are the ones with the most to gain from following GitHub. If the window closes, it will close the way Copilot’s did — in a dated changelog entry, weeks before the renewal that makes you feel it. We log those entries the day they land.

Watch the billing shifts as they land →