The Flat-Rate Era of AI Coding Subscriptions Is Over

For two years, the standard advice for anyone spending serious hours in an AI coding tool was simple: get the subscription, because flat-rate pricing is the best deal a heavy user can get. This page is the counter-case, and it rests on three dated events you can verify yourself — Cursor's June 2025 credit-system overhaul, GitHub Copilot's June 1, 2026 switch to usage-based billing, and Google's October 9, 2026 model-access cut for the Gemini app. Three different vendors, one direction of travel: the more you use, the more the pricing models are being rebuilt to charge you.

The three data points, with dates

Each row below is independently verifiable against the vendor's own documentation or, for the Gemini change, Google's updated help pages.

Vendor	Date	What changed
Cursor (Anysphere)	June 2025	Fixed "fast request" allotments retired; plans became usage-based credit pools — Pro $20/mo now includes roughly $20 of metered usage, with API-key billing beyond it
GitHub Copilot	June 1, 2026	All plans moved from premium-request units to AI Credits under usage-based billing; 1 credit ≈ $0.01 of metered usage
Google Gemini app	October 9, 2026	Free tier cut to Flash-Lite only; the $4.99 AI Plus tier loses Pro model access at an unchanged price

Copilot's credit math is worth stating precisely because secondhand summaries garble it. Per GitHub's own plans documentation, the monthly allowance splits into a base and a flex allotment: Copilot Pro ($10) gets 1,000 + 500 = 1,500 credits, Pro+ ($39) gets 3,900 + 3,100 = 7,000, and Max ($100) gets 10,000 + 10,000 = 20,000. Some third-party write-ups quote only the base numbers, others only the totals — both are quoting the official page, just different columns. At the $0.01-per-credit metering rate, Pro's total allowance is worth about $15 of usage on a $10 plan; Max's is worth about $200 on a $100 plan.

Note the shape of that ladder. The cheap plan's included usage covers about 1.5x its price; the top plan's covers about 2x. Neither is a loss leader, but neither is the unlimited flat rate the subscription model used to advertise. Usage above the allowance bills metered — which is exactly the billing model subscriptions were supposed to protect you from.

Why heavy users specifically

The mainstream framing treats these moves as general price increases. They aren't. A light user on Copilot Pro who burns 400 credits a month is untouched by any of this — the $10 plan got no more expensive, and their allowance covers them with room to spare. What changed is the ceiling for the heavy user.

The reason is mechanical. Subscription pricing in 2023–2025 was priced around chat-shaped usage: a human typing prompts, reading answers. Agent-shaped usage broke that model — an autonomous agent can consume a month of "reasonable" tokens in an afternoon, and the vendors all discovered the same thing at the same time. Cursor's June 2025 change followed user complaints about request-based plans straining under agent workloads; GitHub's own announcement framed usage-based billing as a way to pay for what you actually consume. Translation, read from the vendor side: the heaviest users were unprofitable at flat rates, and rather than raise prices for everyone, the vendors converted heavy usage into a metered line item.

That's a defensible business decision, and this page isn't arguing it's evil. The argument is narrower: if your usage is heavy, the subscription-to-usage conversion changed your effective price, and probably by more than any sticker price change ever did.

The counter-argument, taken seriously

The strongest objection: usage-based billing is fairer, not greedier — light users now subsidize less of the heavy users' burn, and metered pricing scales with value received. If you run agents eight hours a day, maybe you should pay more than someone who codes with Copilot twice a week.

Fair point, with two caveats. First, the included allowances did rise for the top tiers — Max's 20,000 credits is real headroom — so the change isn't purely extractive. Second, "fairer" depends on a number most users have never seen: their own monthly burn. Metered pricing is only fairer if you can predict your bill, and agent workloads are exactly the kind that make bills unpredictable. A model where heavy usage is metered but invisible until the invoice is the worst of both worlds — and that's the one to fix, with budgets, alerts, and a monthly reconciliation habit.

What to actually do at renewal

• Measure one month of real usage before renewing anything. Cursor's dashboard and GitHub's credit ledger both show per-feature consumption. If your burn sits well inside the included allowance, the subscription is still your best deal — this page's premise doesn't apply to you.
• Check what your credits actually buy. Copilot's credit metering varies by model, and Cursor's credit costs differ between its included models and API-billed premium ones. The plan name tells you the allowance; the model picker decides how fast it evaporates. Routing routine work to cheaper models is now a billing decision, not just a quality trade-off.
• Price the metered alternative. If you consistently blow through the allowance, run the same workload through the vendor's metered API (or an equivalent) for a month and compare invoices. For API-shaped work, our Claude API heavy-user cost breakdown shows how caching, batching, and model tiering cut metered bills by multiples — the same discipline applies to any provider's meter.
• Watch the cheap tiers, not just the ones you're on. Gemini's October 9 change hit the $4.99 tier without changing its price. When a vendor devalues the bottom of its ladder, it's telling you how it sees the margin there — and it's the free and entry tiers of every AI tool that will tighten first. We track that move in detail in the Gemini free-tier cut.

The bottom line

The subscription advice of the flat-rate era — "subscribe and stop worrying about tokens" — was good advice for a market that no longer exists. What replaced it is a hybrid: a floor of included usage, a meter above it, and a vendor watching the heavy-user margin on both. The winners in that market aren't the people who pick the right plan once; they're the people who reconcile usage against allowance every month and switch when the math says so. The tools made bills metered. Make your attention metered too.

Sources: GitHub Docs — Plans for GitHub Copilot (AI Credit allowances, retrieved 2026-10-07) · GitHub Blog — Copilot is moving to usage-based billing (Apr 27, 2026) · Cursor — Pricing (plans and credit structure) · Vantage.sh — Cursor Pricing Explained (June 2025 overhaul)