On the morning of Aug. 25, a familiar message returned for some ChatGPT Plus subscribers: usage limit reached. OpenAI has restored its rolling five-hour cap for the standard tier of ChatGPT Plus, ending a six-week experiment with unlimited access that began July 13. The limit applies to users of Codex and ChatGPT Work, the company’s environments for software engineering and professional tasks.
Thibault Sottiaux, OpenAI’s engineering lead for Codex and ChatGPT, confirmed the change on X, saying the rolling five-hour limit took full effect on Aug. 25. Under the policy, a subscriber can use the service for up to five hours of active work within any rolling five-hour window. The cap resets continuously, and it applies per account, not per device.
The return to limits is a quiet admission of the economics of frontier AI. Every conversation with the most advanced models costs OpenAI real money in computing, and the cost scales with how hard the model thinks. Agentic features, which let ChatGPT plan, browse, and write code over long sessions, consume far more compute than a quick question and an answer. Six weeks of unlimited access gave the company a stress test of demand, and the stress, apparently, was real.
The episode fits a pattern. OpenAI has swung between generosity and restraint as it tries to turn ChatGPT from a cultural phenomenon into a durable business. Free tiers get the fastest models for only a few messages a day; paying subscribers expect more, but the company must balance retention against the cost of serving power users. The five-hour cap is a compromise: most subscribers will never hit it, while the heaviest users are nudged toward the company’s more expensive tiers.
The move also reflects the competitive stakes. OpenAI’s rivals, including Google with Gemini and Anthropic with Claude, have been aggressive on pricing and features, and subscription churn in the consumer AI market is high. A cap that frustrates the most engaged users carries risk: those are precisely the users who tweet, recommend the product, and pay for upgrades. OpenAI has tried to cushion the change by grandfathering some entitlements and directing heavy users to enterprise plans, according to people familiar with the company’s rollout plans.
Industry analysts said the decision signals that the cost of agentic AI remains the central constraint on the consumer market. Model makers have talked for years about making intelligence cheap enough to be a utility; the five-hour limit suggests that day has not arrived. Infrastructure costs, including the data centers and power contracts OpenAI has signed, will keep pressure on margins even as revenue grows.
For subscribers, the practical effect depends on how they use the service. Casual users who ask a handful of questions a day will notice nothing. Developers who leave Codex running for hours will feel the limit within days. OpenAI’s documentation advises heavy users to check their usage meters and to consider the company’s higher-tier plans, which carry longer or unlimited windows.
The timing is notable for another reason. The cap returns just as OpenAI is pushing ChatGPT deeper into workplaces, where long, uninterrupted sessions are the norm. If the five-hour limit frustrates exactly the professionals the company is courting, the blowback will arrive in the form of canceled seats and skeptical procurement reviews. OpenAI’s answer, according to people familiar with its strategy, is that limits are a temporary measure, to be relaxed as inference costs fall and new hardware comes online.
The company has been here before. ChatGPT’s popularity in early 2023 crashed OpenAI’s servers within weeks of launch, forcing the company to throttle access and apologize. Each cycle since, from image generation to voice features to agentic coding, has followed the same curve: a feature goes viral, demand outruns supply, and limits return. The five-hour cap fits that history. It also fits the pattern of the broader industry, where every major model maker has quietly reintroduced usage limits after promising abundance. What is unusual this time, analysts note, is the timing: the cap arrives not after a product launch but in the middle of a push to convert corporate customers, which suggests demand for agentic features is running ahead of the infrastructure built to serve it.
The company declined to say how many subscribers the cap would affect, saying only that usage policies are reviewed regularly. Analysts estimate that the heaviest users, a small fraction of the subscriber base, account for a disproportionate share of inference costs, and that concentrating the change on them lets OpenAI preserve the experience of the majority while cutting its own bill.
For now, the return of the cap is the clearest public sign yet of the arithmetic behind conversational AI: every reply has a cost, and someone must pay it. The question is whether that arithmetic, and the limits it imposes, becomes a permanent feature of the AI subscription business. The company that promised to make intelligence broadly available is learning, in public, what that promise costs.
