If you use AI coding tools on a subscription plan, you have almost certainly hit a usage wall. You were in the middle of a productive session, the model was helping you ship real features, and then, suddenly, you are told to come back next week. Welcome to the era of weekly limits.

From rolling windows to hard ceilings

It was not always like this. When Anthropic first launched the Claude Pro plan, usage was governed by a 5-hour rolling window. If you hit your cap, you could grab a coffee, come back, and keep going. It was annoying, but manageable.

Then came the weekly limit. Introduced around August 2025, it added a hard ceiling on top of the rolling window. Once you exhaust your weekly allocation, the 5-hour reset does not save you. You simply wait until the following week. For Pro subscribers paying $20 per month, that means roughly 40 to 80 Claude Code hours per week, and when they are gone, they are gone.

The community reaction was sharp. Anthropic had sent out emails assuring users that fewer than 5% would be affected. But when the limits rolled out, many regular users found themselves hitting 40% of their weekly cap after just a couple of coding sessions. The mismatch between expectations and reality eroded trust.

It is not just Anthropic

What makes this frustrating is that weekly (or monthly) limits have become an industry-wide pattern. Nearly every major AI coding tool has adopted some version of it.

Cursor moved away from its old system of 500 requests per month and now gives Pro users $20 of included usage, measured by actual token consumption. Once you burn through that, you can either pay on-demand rates or upgrade to a higher tier. Power users routinely report $60 to $100 in monthly usage, and some exceed $200.

OpenAI Codex, bundled into ChatGPT subscriptions, follows a similar structure. Plus subscribers ($20 per month) get 30 to 150 messages per 5-hour window, while Pro subscribers ($200 per month) get 300 to 1,500. The range is enormous because a simple script costs far fewer tokens than a complex, context-heavy refactor. Users on the Pro plan have reported hitting their 5-hour limit within two hours and nearly exhausting their weekly cap in a single day.

T3 Chat, built by Theo from T3.gg, ran into the same economics from the other side of the table. At $8 per month for 1,500 messages, the math worked fine for most users. But roughly 2% of subscribers were responsible for the vast majority of costs, with individual users racking up over $200 in compute against an $8 subscription. Theo was candid about it: "This is not a change I wanted to make." T3 Chat has since overhauled its credit system entirely, replacing monthly message counts with a usage bar that resets every 4 hours.

Why the limits exist

The economics are straightforward, even if the experience is not. Running frontier AI models is expensive. GPUs cost real money, electricity bills are enormous, and inference at scale is not cheap.

Subscription pricing was designed to be generous. $20 per month for near-unlimited access to a model that would cost significantly more through the API is, in many cases, a losing proposition for the provider. Anthropic, OpenAI, and others are effectively subsidizing usage to build market share and gather feedback. The limits exist because, without them, a small percentage of heavy users would make the entire model unsustainable.

The numbers are stark. In March 2026, Bearly AI surfaced Cursor's internal analysis showing just how aggressively Anthropic is subsidizing Claude Code. Last year, a $200 monthly Max subscription could consume roughly $2,000 in compute. Now, that same $200 plan can burn through $5,000, a 2.5x increase. The subscription price stayed flat while the compute consumed per user ballooned.

This creates a two-tier economy where developers and companies paying API rates are effectively funding the infrastructure that subscription users access at a steep discount. Theo explored this dynamic in his video "I was wrong about AI costs (they keep going up)," drawing on Ethan Ding's analysis of what Ding calls the "token short squeeze." The thesis is counterintuitive: despite per-token costs falling roughly 10x per year, the actual cost of serving users keeps climbing. The frontier models that everyone wants never get cheaper. What gets cheaper is last year's model, which nobody wants. Meanwhile, reasoning models, agents, and long-running tasks have caused token consumption to explode. A single deep research run can cost over a dollar. Wes Bos shared that he burned through $118 worth of tokens in a single day, and noted that many developers routinely hit $350, all at already-subsidized rates. API customers paying full price are, in effect, keeping the subscription model alive.

From the provider's perspective, this still makes sense. The subscription is dramatically cheaper than paying API rates for the same volume of work. A developer who uses Claude Code heavily through the API could easily spend hundreds of dollars in a single week. At $20 per month, even with weekly caps, the value proposition is objectively strong for the user. But the gap between what subscribers pay and what they consume is widening, not shrinking, and it is API revenue and venture capital that fill the hole.

The frustration is still real

Understanding the economics does not make the experience less annoying. The core tension is this: these tools are genuinely useful. They change how you write code, debug problems, and explore unfamiliar codebases. When you are in a flow state and the tool cuts you off, the disruption is disproportionate to the cost savings.

There is also a psychological dimension. Theo from T3 Chat observed that even when 99.9% of users never come close to hitting their limit, the mere existence of a cap discourages people from subscribing. The limit becomes a mental barrier before it becomes a practical one.

For heavy users, the options are not great. Claude offers Max plans at $100 (5x usage) and $200 (20x usage) per month. Cursor lets you add on-demand usage at API rates. But for an individual developer, especially outside the US where exchange rates make dollar-denominated subscriptions more expensive, these tiers can feel out of reach.

This is not going away

If anything, weekly and monthly limits are becoming more common, not less. As AI models get more capable and context windows get longer, the cost per session increases. Providers who launched with generous allowances are tightening them. New entrants are building limits into their pricing from day one.

The pattern is clear: subscription plans will remain the most cost-effective way to use AI coding tools, but "unlimited" is a word you will see less and less. Rolling windows, weekly caps, token-based metering, and tiered plans are the new normal.

For users, the practical advice is simple. Be intentional about how you use your allocation. Save complex, context-heavy tasks for when they matter most. Use lighter models for simpler queries. And accept that, for now, the tools are good enough that even with limits, they are worth paying for.

That might be the most telling sign of all. Despite the weekly caps, the rate limit warnings, and the mid-session cutoffs, people keep subscribing. The tools are that useful. The limits are the price of admission.

References