For years, the flagship AI subscription worked like a gym membership: pay a flat $20 a month, show up as often as you like, and let the provider eat the variance. That era is ending. This month, Anthropic trimmed Claude Code's weekly limits while announcing an "increase." In June, Google rebuilt Gemini's subscription around a compute-based meter. GitHub moved Copilot to usage-based AI Credits, and OpenAI re-priced Codex by token. The sticker price on your plan may not have changed — but what it buys you has. Your subscription is quietly becoming a compute budget.

The increase that was a cut: Claude Code's September 14 change#

In late August, Anthropic announced a permanent 25% increase to Claude Code's standard weekly limits, effective September 14, 2026, covering Pro, Max, Team, and seat-based Enterprise plans. Five-hour session limits were unaffected. Framed as an upgrade, it landed badly — because developers did the arithmetic within hours.

Since May 13, Anthropic had been running a temporary 50% boost to weekly allowances, extended four times over the summer. Index the original baseline at 100: users have been living at 150 for months. The "permanent" level is 125 — 25% above the old baseline, yes, but about 17% less than what subscribers actually use today. A Community Note appeared under the original announcement spelling out the math, and Anthropic followed up with a clarification admitting the figure outright: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code."

Both numbers are accurate. Which one you lead with decides whether this reads as an expansion or a squeeze — and for anyone who built a working week around the 150 level, it is a squeeze. Anthropic said the pullback was needed for long-term platform stability and responsible compute management. It also teased better visibility: dashboard tools promising clearer views and more direct control over remaining weekly quotas.

The episode matters less for its 17% than for what it reveals. Heavy users learned that their allowance is not a fixed product feature — it is a number the vendor re-tunes against GPU supply, and promotional headroom eventually gets reclassified as normal, then reduced. The same article that reported the change noted that OpenAI recently restored strict five-hour limits on its own Codex tools. GPU-heavy workloads are straining server capacity everywhere, and subscribers are absorbing the adjustment.

Gemini's quiet pivot to compute-based metering#

Google made a similar move earlier and more explicitly. In June 2026, Google updated its AI subscription plans and introduced what it described as a new way of managing usage limits inside the Gemini app: a compute-based model that weighs prompt complexity, the features being used, and overall conversation length. Limits refresh every five hours up to a set weekly cap — a structure that now looks nearly identical to Claude's session-and-weekly model. AI Plus subscribers get roughly double the free tier's limits, and Google began phasing out its old 200 monthly AI credits in the process.

The pricing rungs tell the rest of the story. Google AI Pro has held at $19.99 a month; AI Ultra was split at Google I/O 2026 into a $99.99 tier and a $199.99 tier with up to 20x Pro's limits. The gradient from $0 to $200 is no longer about feature gates — storage, Deep Think, YouTube Premium — it is mostly about how much compute you may spend. Even Google's own tracker pages now describe the tiers in terms of limit multipliers (Pro at 4x free, Ultra at 5x–20x Pro), and the Gemini CLI docs publish hard per-day request quotas: 1,000 for Code Assist Individual, 1,500 for AI Pro, 2,000 for AI Ultra, with pay-as-you-go available when you outrun the cap. The meter is published; the flat plan is the on-ramp.

The wider meter: Copilot, Codex, Cursor#

Anthropic and Google are not outliers — they are the last to admit what the rest of the market already did:

  • GitHub Copilot — Effective June 1, 2026, every Copilot plan shifted to usage-based "AI Credits," consumed at published token rates for agentic work while simple completions stayed unmetered. GitHub's CPO was blunt about the old model's death: "A quick chat question and a multi-hour autonomous coding session can cost the user the same amount… the current premium request model is no longer sustainable."
  • OpenAI Codex — On April 2, 2026, OpenAI moved Codex from per-message allowances to token-aligned pricing. Codex remains included across ChatGPT plans, but the allowance is now measured in tokens, with optional credit packs for power users who hit the ceiling. CEO Sam Altman described the philosophy directly: keep base subscriptions cheap and predictable for the majority, and let a small set of power users pay to go past the caps.
  • Cursor — the early mover, and the cautionary tale. It pivoted to credits priced at underlying API rates in mid-2025 and fumbled the transition badly enough — one reported case of a $7,000 annual subscription consumed in a single day — to force a public apology. The message users heard was not "we raised prices" but "your workflow is now a metered compute budget."

The pattern is consistent: the quarter where the meter switched on was the second quarter of 2026. Flat-rate AI subscriptions were a customer-acquisition tactic for the autocomplete era. Agentic tools are compute products. Compute products revert to metering.

Why flat-rate couldn't survive agents#

The economics are straightforward. A chat question costs fractions of a cent; an autonomous coding session — dozens of tool calls, repo-wide context, long reasoning chains — can cost orders of magnitude more. Under flat pricing, one user's afternoon of agentic work subsidizes another's week of light chat, and the power users are exactly the customers the product is built for. Variance killed the gym membership.

There is also a subtler dynamic worth naming: quotas don't translate into finished work. Because conversation length, model selection, and tool execution change context size on the fly, a weekly allowance can't be read as a simple prompt count. One complex task can burn half a five-hour window. That unpredictability is why the backlash hits so hard — and why every vendor teasing a limit change also teases a usage dashboard. Metering without visibility is just a surprise bill.

Your new compute-budget playbook#

Whether you code, research, or run an agency on these tools, the posture has to change from "subscribe and forget" to "budget and measure." A few practical moves:

  • Re-measure against the real number, not the promo number. Claude Code users should size their week at ~125, not 150, from September 14. If the reduced limit doesn't fit your workflow, compare plans or providers on current terms — promotional headroom is not a plan.
  • Route work to the right model. Reserve frontier reasoning models for genuinely hard problems; use cheaper, faster models for formatting, simple edits, and boilerplate. The per-token cost difference between model tiers is often 5–10x.
  • Watch the dashboard, not the price. The new usage panels Anthropic and others are shipping aren't decoration — they are the prerequisite for living under metering. Check utilization before starting a big agentic session, not after you hit the wall.
  • Keep switching costs low. Avoid hard-locking your workflow to one vendor's proprietary features. Multi-provider setups — routing through a gateway or simply knowing your fallback — turn one vendor's limit change from a crisis into a config tweak.
  • Set caps for teams, not just subscriptions. If you manage developers on metered tools, the team-level bill can jump far faster than seat count suggests. Budget per workload, not per head.

Takeaway#

The unlimited AI subscription was a limited-time offer from the venture-funded land-grab years, and the fine print is being rewritten in public. Claude Code's 17% cut, Gemini's compute-based meter, Copilot's credits, and Codex's token pricing all point the same direction: the plan you pay for is now an allowance of compute, refreshed on a schedule, priced against real GPU cost. The labs are not hiding this — they are publishing the meters and shipping the dashboards. The users who thrive under the new regime will be the ones who treat their subscription the way they've always treated cloud infrastructure: measured, budgeted, and routed with intent. The meter is running either way. You might as well be reading it.