Claude Code Usage Limits Drop 17% on Sept 14: Inside the GPT-6 Astra Reset War

Anthropic's Claude Code subscription becomes meaningfully less generous on September 14, when a temporary +50% weekly-usage boost that has been running since May is replaced by a permanent +25% increase — which, measured against what subscribers get today, is really a cut of roughly 17%. The move is one round in the strangest competitive fight in AI right now: a quota-reset war between Anthropic, whose Claude Code and Claude chat subscription pool draws on its flagship Claude Fable 5.1, and OpenAI, whose Codex product is built around GPT-6 Astra. It is a fight conducted with usage resets and product-manager theatrics, and it is now changing what a paid subscription is actually worth from week to week.

On September 5 the developer commentator kimmonismus posted a widely shared summary of the standoff: "this is competition at its best… the release of Tibo has forced Anthropic to finally reset the rates, something they would never have done otherwise. At the same time, Anthropic has forced OpenAI to refocus on what's essential: model capability." "Tibo" is Thibault Sottiaux, OpenAI's product lead for Codex and ChatGPT, who has spent the summer resetting paid users' quotas so often that the community nicknamed him the reset god. The joke in the post — the "I smell fear" meme — is Tibo's own, from a July night when he reset Codex limits minutes after Anthropic reset Claude's. What follows is not about benchmark scores. It is about the other half of the model economy: how much work a flat monthly fee actually buys, and who decides when that number changes.

What actually changes on September 14

The Claude change is real and datable. On August 29, Anthropic's ClaudeDevs account announced that from September 14, standard weekly Claude Code limits will be permanently raised by 25% for Pro, Max, Team, and seat-based Enterprise plans. What the announcement did not put first is that subscribers are currently running on a temporary +50% weekly boost that has been in place since May 13 and runs through September 13. Replacing a +50% bonus with a +25% bonus is a step down, and the arithmetic is straightforward on Anthropic's own baseline:

• Today: baseline 100 + 50% boost = 150 units of weekly quota.

• September 14: baseline 100 + 25% permanent increase = 125 units.

• Net change for existing users: roughly −17% (150 to 125).

Anthropic later acknowledged the math directly — "compared to today, this works out to a 17% reduction" — after the discrepancy drew community notes and developer criticism. The company framed the change as ending a temporary expansion while making part of it permanent, pointed to long-term sustainability, and said it is working on "exciting changes" to give users more visibility and control over usage. It announced no dates for those features.

Two structural facts make the cut harder to shrug off. First, on Pro and Max plans, Claude Code shares a single usage pool with Claude chat and Cowork, so the same weekly allowance is spent across three products. Second, none of it is expressed in tokens: Anthropic describes limits as multipliers rather than budgets, so the way users learn they have hit a ceiling is by hitting it. A step-down announced as a raise lands differently when you cannot see the underlying number.

The other side of the war: OpenAI keeps handing out resets

OpenAI has spent the same summer moving in the opposite direction, and doing it loudly. Sottiaux has distributed more than 40 all-user Codex reset coupons since the spring — the community tracks them at codex-reset.com — most valid for 30 days. A "banked reset" is a one-time stored refresh of both the 5-hour and weekly usage windows that the user chooses when to spend; it cannot be bought, cashed out, or transferred, and it expires. In dollar terms the resets are meaningful: paid users have tested them at the equivalent of roughly $5–8 of usage on Plus and $50–80 on Pro, while community tracking (NerfTrack) estimates a Plus weekly allowance is worth on the order of $80–160 of GPT-5.6 Sol usage — and heavy users get far more than the median. One Pro user was tracked consuming 2.5 billion tokens in 23 hours, about 80% of a weekly allowance.

On August 30, Tibo announced a full reset for all paid Codex and ChatGPT Work users — and, unusually, explained why their quotas kept draining. He said the team had fixed a set of bugs that were silently eating usage: image compression and compaction, background memory workers, /goal retries, sub-agents, computer-history overhead, and MCP-related background work. Depending on usage patterns, the same allowance should now last 10% to 50% longer. That itemized disclosure is the thing Anthropic has not matched: a list of where the tokens actually went.

The latest reset came this week. After GPT-6 Astra finished rolling out to paid ChatGPT tiers — a rollout that was rough at the start, with a ChatGPT and Codex outage on September 3 and the model initially inaccessible even to some $200-per-month Pro subscribers — Sottiaux announced a global reset of usage for all paid ChatGPT subscriptions, timed for September 7 around 6 pm Pacific (the morning of September 8 in Beijing). The announcement was itself a bit: a Rickroll post whose emphasized lyrics spelled out RESET. The stated reason was to give users who had maxed out during the Astra surge a clean window for the start of the work week.

Why the two companies are now fighting over quotas

None of this looks like the benchmark war of a year ago, because the models have converged on price. Claude Fable 5.1 — released by Anthropic on September 1 as the public, deployed version of its Mythos-class tier, at $10 per million input tokens and $50 per million output — and GPT-6 Astra, OpenAI's flagship from September 3 at the same $10/$50 list price, are close enough on capability that scoreboards no longer settle the argument. Claude Fable 5.1 tops the independent Artificial Analysis Intelligence Index at 66, with GPT-6 Astra at 61, and Anthropic's launch also cut cache-read pricing 75% to $0.25 per million tokens, framing itself as the cheaper option for long-running agents.

When models tie on price, the fight moves to seats. Both companies now sell heavy users the same thing: an agentic coding subscription with a usage cap neither will state in tokens. OpenAI's lever is reset generosity; Anthropic's, until this month, was capacity — the temporary boosts trace to Anthropic's May data-center expansion. That is why a quota reset now reads as a competitive event and a quota step-down reads as retreat, and why the two product leads perform for each other. Tibo's "I smell fear" reply in July — after the GPT-5.6 Sol launch pushed Anthropic into a same-day Claude reset — set the tone, and the Claude side now has its own reset-happy figurehead: reports this week describe Anthropic quietly resetting Claude weekly limits again under GPT-6 Astra pressure, with Lydia Hallie playing the counterpart the community has taken to calling the cyber godmother. The kimmonismus post is right that users benefit — the quarrel has produced dozens of free quota resets — but it has also turned subscription terms into a marketing instrument, which is exactly when terms stop being predictable.

What a metered API does that a subscription cannot

The practical question for anyone paying for one of these subscriptions is whether the gamesmanship should change what you run where. A weekly allowance that silently steps down 17%, or resets on a competitor's schedule, is a planning problem: you discover your ceiling by hitting it, and neither company will print the token budget. That uncertainty is the strongest argument for keeping a metered path in your stack. A subscription is a bet on how much work a flat fee buys; an API is a price per token actually consumed. On OrcaRouter, both models at the center of this fight are available at their provider list prices with no markup — Claude Fable 5.1 at $10/$50 per million tokens with a $0.25 cache-read price, and GPT-6 Astra at $10/$50 with its $20/$75 long-context tier above 272K tokens of input. Because the platform passes the provider's list price through at zero markup, Anthropic's cache-read cut was live on our side the same day it was announced — no renegotiation and no second contract when one vendor blinks first.

Routing also changes the risk calculus of the war. When Codex went down on September 3, a workload routed with a failover rule did not stop; the request went to an alternative model automatically. When you want to know whether Claude Fable 5.1's agentic gains justify its price for your specific codebase, you can send a slice of traffic to it and the rest to GPT-6 Astra through one API key and compare cost per completed task — no subscription to buy or cancel, no banked reset to time. The quota war is a feature of subscriptions; on a metered API, one vendor's reset politics changes nothing about your access.

What to watch between now and September 14

• September 14. Whether Anthropic holds the line on the step-down or extends the +50% boost again under pressure — it has already extended the boost several times this summer, and reports this week say Claude has begun resetting weekly limits again.

• Whether OpenAI's resets stop when the rollout ends. Tibo has said he will keep resetting until GPT-6 Astra is genuinely available across ChatGPT tiers; once Astra is fully live everywhere, the pressure valve closes and the resets are likely to become less theatrical.

• Anthropic's promised usage visibility. The company says it is working on giving users more control and visibility over usage; if it ships an actual token or dollar readout, it would be the first of the two to stop hiding the budget.

• A drain-fix disclosure from Anthropic. OpenAI itemized what was eating Codex quotas; Claude users are still guessing at what background memory and compaction cost them.

• The model front, where the war may escalate anyway. An unverified September 5 leak claims Anthropic is already readying a successor in the Claude Fable 5.2 line as its answer to GPT-6 Astra — the same competitive pressure that produced this week's quota maneuvering, aimed now at the models themselves.

The reset war has been good for users in the short run — the past week alone brought a global Codex reset and renewed Claude resets, and the two companies are now visibly bidding for heavy users in a way that benefits everyone who pays. But generosity that a product manager can switch off is not a pricing model, and a −17% step-down dressed as a +25% increase is the clearest reminder yet that subscription quota is a marketing instrument, not a contract. The models underneath — Claude Fable 5.1 and GPT-6 Astra — are excellent, closely matched, and identically priced; the terms around them are the part that keeps moving. Meter what you can, watch September 14, and do not let a reset schedule decide your architecture.