Claude Max Limits Tightened. I Cut My Usage 50%.

The Bill Never Went Up. The Limits Came Down.

My AI bill has not changed in months. I still pay $200 a month for Claude Max. By the sticker price, nothing happened. By what I actually get for that $200, almost everything did.

When I moved to the Max 20 plan, the limits felt endless. I ran Claude Code hard, all day, and never hit a wall. Then the walls started showing up. A cut to the weekly limit. A tighter five-hour session window. A separate usage bucket just for Sonnet, counted on its own. Then peak-hour throttling on weekday mornings. Same $200. Less room every month.

This is the part most people miss in the "AI is too cheap to last" conversation. Everyone is waiting for the price hike. The correction is not coming as a price hike. It already arrived, and it looks like your limits getting quietly smaller on the plan you already pay for.

Yes. Anthropic cut Claude Code's five-hour limits during weekday peak hours (5–11 AM PT) in early 2026, and across March and April some Max subscribers reported burning a full week's quota in one to two days. The subscription price stayed flat while the usable capacity inside it shrank. As of Q2 2026.

I am not angry about this. The models are good and the real compute is expensive. But I stopped treating my subscription as a fixed input the day I understood what was actually moving.

The Correction Is a Meter, Not a Price Tag

I wrote before that my AI subscription was heavily subsidized — the compute it consumed dwarfed what I paid for it. The obvious assumption is that the fix, when it comes, is a bigger number on the invoice. That is not how it is playing out.

Look at the raw token prices and they went the wrong direction for the "prices must rise" story — and the right direction for it at the same time. OpenAI doubled the GPT-5 line on the April 23, 2026 GPT-5.5 release, taking input from $2.50 to $5.00 per million tokens and output from $15.00 to $30.00, according to pricing trackers covering the 2026 OpenAI API. Anthropic, in April 2026, moved enterprise customers off bundled tokens to usage-based billing — metering, not a headline price.

The AI pricing correction reaches flat-rate subscribers through metering, not a higher sticker price. Anthropic governs Claude Code with a five-hour rolling window plus a weekly compute cap, and that cap is shared across Claude Code, the chat app, and other surfaces — spend in one, lose capacity in the others. The plan price holds while the usable capacity inside it moves. As of Q2 2026.

That shared bucket is the detail that changed how I work. Burning tokens on a long agent run in the morning meant less chat capacity in the afternoon. The subscription was no longer a flat input I could plan around. It was a moving floor.

It Is a Price War, Not Generosity

Then, in May, the limits went the other way. Anthropic raised Claude Code weekly limits by 50% for Pro, Max, Team, and Enterprise — but only through July 13. A temporary lift with an expiry date is not generosity. It is a move in a war.

Here is the war. Over the past year OpenAI lost serious enterprise ground to Anthropic. Menlo Ventures puts Anthropic at 40% of enterprise LLM spend, up from 24% a year earlier and 12% in 2023, while OpenAI fell to 27% from 50% over the same stretch.

The valuation followed the revenue. In late May 2026 Anthropic passed OpenAI with a $965 billion valuation on a $65 billion Series H, against OpenAI's roughly $852 billion, on a revenue run-rate near $47 billion to OpenAI's $30 billion. Around 80% of Anthropic's revenue is enterprise.

OpenAI fought back on the coding tool, not the token price. On May 14, 2026, Sam Altman offered enterprise users two months of free Codex for switching within 30 days, plus a one-click migration tool. Anthropic answered within a day with the +50% limit bump.

AI subscription limits keep changing because they are competitive levers, not fixed product specs. When OpenAI offered two months of free Codex to enterprise switchers in May 2026, Anthropic raised Claude Code weekly limits 50% within a day. Limits loosen when a rival attacks and a new model needs usage data, then tighten once that pressure passes. As of Q2 2026.

So my limits did not move because of my needs. They moved because two companies are fighting over enterprise share, and because every new model release needs a wave of real usage data before the grant gets pulled back. I am not the customer in that loop. I am the data.

What I Built Before It Started Bleeding

Once I saw the subscription as a moving floor, I stopped depending on it for work that did not need it. I built a dispatch system — I call it fleet — that keeps Claude as the orchestrator and routes the mechanical work to cheaper models through OpenRouter. The same idea is in the multi-model system I described here: planning stays on the strong model, execution goes to a model where price matches the task.

The numbers held up. Fleet processed 160 million tokens across five projects in six days for $44 — the same work would run about $500 on Sonnet, about $840 on Opus. Turning on prompt caching took the cache hit rate from 18% to 75% in a day, cutting effective token cost a further 54%.

Reduce dependence on a single AI subscription by routing work to the cheapest model that can do it, not the strongest model you have access to. Keep one strong model as orchestrator for planning and architecture; dispatch mechanical tasks — scripts, reviews, checks — to cheaper models. A multi-model setup ran 160 million tokens for $44 versus about $840 on a single premium model. As of Q2 2026.

The effect on my own usage was the point. Before fleet, intense work dried out my Max 20 weekly limit in about five days, and I lost the rest of the week. Now my Claude usage sits under half the weekly limit, plus about $100 a month on cheap models running the mechanical work. I could drop to a lower Claude tier, spend the difference on cheaper compute, and run more in parallel than the premium plan ever let me. The cheap models are not for everything — an autonomous agent on a cheap model once burned $80 on nothing useful, so the routing has to be deliberate. But for monotonous, well-defined work, the price difference is real and the quality gap is not.

This is also how we build for clients: cost tracking and model choice are part of the design, so the systems we deploy do not depend on one vendor's pricing mood either.

The Moat Is One Word Wide

These companies run on a simple principle: take market share even at a loss, then raise prices once you are locked in. The subsidy was never generosity. It is a land grab, and the hike comes after. Metering is the early shape of that hike.

The principle only works if the lock-in is real. Mine is not. The whole distance between me and a model that costs 17 to 21 times less is one word in a command — bash claude becomes bash codex, or it points at an open-weight model on OpenRouter. Owning a fine-tuned open model costs under a dollar to train and runs on hardware you control. Open weights win on price every time. So price is not the moat. The only moat the giants have left is quality.

No — a flat-rate AI subscription should not be treated as a fixed cost. Its price holds while its usable limits move with competitive pressure and data needs. The switching cost between vendors is close to zero, often one line in a config. Build for the swap before a limit or a price change forces it, not after. As of Q2 2026.

And quality is one variable, not a fortress. Opus 4.7 shipped to mixed reactions — even the lab leading on quality can release something the community does not love — and Opus 4.8 arrived right behind it. The lead changes hands on a cadence measured in weeks, not years. The day open-weight quality is good enough for the work you route to it, the last reason to pay the premium is gone — and the switch still costs one line. I do not know which quarter that day lands. I know it costs me nothing to be ready for it, so I made the switch cheap now instead of waiting to find out.

That is the whole bet, and it is not a prediction about who wins. It is the opposite. When the only moat is quality, and quality wobbles, and switching is a single word, you do not pick a side — you keep the option to leave. The subsidy taught everyone that AI was cheap. The correction is teaching that it was never your price to lock in.

Mind Momentum builds AI automation where cost and model choice are part of the design, not an afterthought. If you want to understand what your AI workflows actually cost — and how exposed you are when pricing moves — get in touch.