Fundamentals

Claude Usage Limits: The Commander Fix Nobody Talks About

Burning through Claude usage limits? Stop running everything on the biggest model. Set one commander, let cheaper models execute, and make your plan last.

By Alex Arhontoulis · July 21, 2026 · 11 min read

If you are on a Claude Max plan and still burning through your quota before lunch, the problem is almost never the plan itself. The problem is that you are running every single task on the most expensive model, which is like hiring a senior partner to staple documents. Here is the fix nobody in the complaint threads is talking about.

Why your Claude usage disappears so fast

Most people hit the wall, go to Reddit or GitHub, read that there is a rolling 5-hour window and a weekly cap, and assume the answer is to upgrade. Sometimes upgrading is right. But most of the time the real issue is sitting one layer deeper, in how you are using the models, not which tier your credit card is paying for.

The two limits stacked on top of each other

Claude Max plans run two separate limit clocks at the same time. There is a rolling 5-hour usage window that resets continuously, and there is a separate weekly cap sitting on top of that. Most power users I talk to know about one of these but not both. You can burn the 5-hour window in under two hours on heavy tasks, think it reset, and then smack into the weekly ceiling before Friday. That double-stack is why it feels like the quota evaporates for no reason. It is not random. You are just getting hit by two different counters.

On Max plans there is also a separate, more restricted limit specifically on Sonnet-tier usage in certain interfaces. So even if you are trying to be smart about not running everything on the biggest model, you may find Sonnet has its own ceiling to watch.

One pool shared across Claude Code, claude.ai, and Cowork

Here is the part that surprises people. Your usage quota is not split by surface. Claude Code, claude.ai in the browser, and Cowork all draw from the same pool. So if you spend a big chunk of your window in Claude Code debugging a script, you come back to claude.ai and wonder why it feels sluggish, that is why. There is one bucket and every tool you use dips into it. Track where you are spending before you assume you have room.

The real culprit: every task running on the biggest model

This is the one nobody talks about. The newest top model, whether that is Fable 5 or Opus, is designed to think longer per task by design. That is a feature when you need judgment. It is pure waste when you are asking it to rename a list of files or format a table. The same number of tasks run on the biggest model can cost 2x to 10x what they would cost on a cheaper executor. Your quota disappears because the meter is running at the top rate even for mechanical work that a smaller model handles just as well.

This week, pull up your last 10 tasks and ask honestly: did any of those actually need the smartest model in the room?

The price of using a sledgehammer for everything

Let me put actual numbers on this so it is not abstract. Claude pricing is public and the difference between tiers is significant.

What each model actually costs per million tokens

Model Input (per million tokens) Output (per million tokens)
Claude Fable 5 ~$10 ~$50
Claude Opus ~$5 ~$25
Claude Sonnet ~$3 ~$15
Claude Haiku ~$1 ~$5

These are approximate list prices as of publication. The point is not the exact cents. The point is the ratio. Output on Fable 5 costs 10x what Haiku costs. So if you are doing mechanical output work on Fable 5 that Haiku could handle, you are spending 10 dollars where you could be spending one.

The 2x to 10x waste on mechanical work

Most of what people actually do with Claude is not hard judgment work. It is reformatting, summarizing, extracting data, writing routine replies, drafting first passes of things. That work sits comfortably in the Sonnet or Haiku range. Running it on Fable 5 or Opus is roughly 2x to 10x more expensive per unit of output than it needs to be. If your plan is burning out in a single morning, it is very likely because the mechanical half of your workload is running at the top-tier rate. Fixing that ratio is the fix.

Next step this week: identify two or three recurring task types you run on Claude and decide honestly which tier of model they actually need.

The commander fix

The pattern I use, and the one Anthropic's own developer guidance points toward, is simple. One smart model is the brain of the operation. It plans, reviews the hard calls, and decides what to do next. Cheaper models do the actual work as subagents. The expensive model runs as little as possible. The cheap models run as much as needed.

One smart model runs the show

The commander model handles things where being wrong is expensive. Architectural decisions. Ambiguous instructions. Long-horizon planning. Anything where the model needs real judgment, not just the ability to execute a clear instruction. You point Fable 5 or Opus at these moments and nowhere else.

Cheaper models do the legwork as subagents

Once the commander has a plan, it hands tasks off to cheaper Claude subagents. Those subagents do the formatting, the file work, the drafting, the extraction. They run at Sonnet or Haiku rates. They do not need to plan. They just need to execute a clear instruction well, and they do. This is not a workaround. It is the intended pattern. Anthropic's developer documentation specifically recommends spawning subagents with cheaper models for sub-tasks while keeping the main planning loop on one model. It also recommends low effort settings for subagents handling simple tasks, which cuts token spend further.

Why this is exactly what Anthropic recommends

The newest top models are built to reason longer before responding. That extended thinking is expensive on purpose, because it produces better judgment. But extended thinking on a task that does not require judgment is just wasted tokens. Anthropic's guidance explicitly acknowledges this and recommends matching model capability to task complexity. The commander pattern is not a hack. It is how the system is supposed to be used.

Set up the commander pattern once this week on whatever your most frequent multi-step task is, even a simple two-model split will show you the difference immediately.

How I run it day to day

Here is exactly how I do this at Apex Solved, not as a theory but as a daily practice.

Fable or Opus as the brain for planning and hard calls

When I am scoping a new bot build for a client or making a call about how a system should be structured, that is where Fable 5 or Opus earns its rate. I ask it to lay out the plan, flag the decisions, and identify what can be handed off. It runs for a few minutes, outputs a clear structure, and then steps back. I do not keep it in the loop for execution work.

Sonnet for the building

Once I know what I am building, Sonnet does the actual construction. Writing scripts, building prompt chains, drafting client-facing copy. Sonnet is fast, accurate on instruction-following tasks, and runs at a fraction of the Fable 5 output rate. The quality difference on well-defined execution tasks is minimal. The cost difference is significant.

Haiku for the grunt work

Repetitive output, routine answers, simple extraction, anything that is high-volume and low-complexity goes to Haiku. The AI chat that answers questions on my own website at apex-solved.com runs on Haiku. It costs me pennies per day. It handles the common questions well. I do not need Fable 5 to tell someone what my turnaround time is.

A real example from my own business

I worked with a business owner who was on a Max plan and consistently running out of quota before the afternoon. They were running every single step, planning, drafting, reformatting, sending, through the biggest available model because they assumed bigger meant better. It did not. We split the work into a simple two-tier setup: the top model for planning and review, Sonnet for everything else. The same workload that was blowing the plan by noon fit inside the window with room left. Nothing about their output quality changed. The only thing that changed was which model was doing which kind of work. That is the whole fix.

This week, try running one full task end-to-end using the commander pattern, even manually, just to see how much of it actually needed the top model.

Three more burn-savers

The commander pattern is the biggest lever, but there are three other things that have meaningfully reduced my token spend and will do the same for you.

Turn the effort dial down for simple tasks

Claude has effort level settings, low, medium, and high, that affect how much the model thinks before responding. High effort is appropriate for hard problems. For simple tasks, low effort is dramatically cheaper on token spend and the output is basically identical. Most people leave this on the default or assume higher is always better. It is not. For a subagent that is doing file cleanup or reformatting, low effort is the right setting.

Stop re-pasting giant files into fresh chats

Every time you paste a large document or a long file into a new chat session, you are paying to process that context again from scratch. Repeated context is repeated cost. Keep one working session going as long as the task is live. If you need to start a new session, use references or summaries instead of re-pasting the whole thing. This one habit alone can cut a meaningful chunk out of your usage, especially if you are working with long documents or code files regularly.

Match the model to the task, not your ego

I say this plainly because I have done it myself. There is a tendency to reach for the biggest model because it feels like you are taking the task seriously. It is not a better approach. It is just a more expensive one. The right model for a task is the cheapest one that handles it reliably. If Sonnet gets the job done, using Fable 5 instead is not diligence, it is waste. Match the tool to the job.

Pick one of these three this week and apply it to your next five Claude sessions to get a baseline on how much it moves the needle.

When the big model IS worth it

I want to be straight with you here because I am not trying to talk you into using the cheapest model for everything. There are real cases where Fable 5 or Opus is the right call and using anything cheaper would cost you more in time or mistakes than the tokens are worth.

Use the top model when:

And yes, if your actual volume is high enough that even an optimized model mix is not enough, upgrading your plan is sometimes just the right answer. I am not here to tell you the Max plan is not worth it. For people who use Claude heavily every day as a core work tool, it often is. The point is that most people hitting the wall are not hitting it because the plan is too small. They are hitting it because the model mix is wrong. Fix the mix first. Then decide if you still need to upgrade.

If you are building AI systems that run on behalf of your business, like the custom bots I build for clients through Apex Autobots, getting the model mix right at the system design stage is part of the build. It is not an afterthought. A bot that runs every step on the top model is going to be 5x to 10x more expensive to operate than one designed with the commander pattern from the start.

What to do this week

Here is the practical version, three steps, nothing complicated.

  1. Audit your last 10 tasks. Write down which model ran each one. Then mark each task as either judgment work or execution work. You will almost certainly find that a large portion of your execution work ran on a top-tier model.
  2. Set up the commander pattern for your most frequent multi-step task. Pick the task you run most often. Assign Fable 5 or Opus to the planning step only. Move the execution steps to Sonnet or Haiku. Run it once and see how the output compares.
  3. Apply one of the three burn-savers. Pick effort level tuning, session continuity, or model-to-task matching. Apply it consistently for one week and watch your quota usage.

If you want to talk through how to set this up properly for your business, or if you are interested in having a well-built AI system running for you from day one with the model mix already dialed in, reach out directly. Email me at aaarhontoulis@gmail.com or call (484) 602-6390. I build and deploy these systems for small business owners in 7 days. You bring the problem, I build the solution.

Key takeaways

Common questions before you build.

Why does Claude hit usage limits so fast?

Claude Max plans run two separate limits at the same time: a rolling 5-hour usage window and a weekly cap. On top of that, your quota is shared across every Claude surface you use, including Claude Code, claude.ai, and Cowork. Most users burn through their quota fast because they are running all tasks on the most expensive model, which consumes tokens at 2x to 10x the rate of cheaper models for the same mechanical work.

How do I make my Claude plan last longer?

The single biggest change you can make is stopping the habit of running every task on the top-tier model. Use a smart model like Fable 5 or Opus only for planning and judgment calls, then move execution work to Sonnet or Haiku as subagents. Three supporting habits also help: set effort levels to low for simple tasks, keep one working session instead of re-pasting large files repeatedly, and always pick the model based on what the task actually requires.

What is the difference between Claude models?

Claude models sit on a spectrum from Haiku at the low end to Fable 5 at the top. Haiku is fast and cheap at approximately $1 per million input tokens and $5 per million output tokens, designed for high-volume simple tasks. Sonnet sits in the middle at roughly $3 in and $15 out, strong for most building and execution work. Fable 5, the newest top model, runs about $10 in and $50 out and is designed to reason longer per task, which makes it right for hard judgment calls but expensive for routine work.

What is Claude Fable 5?

Claude Fable 5 is Anthropic's newest flagship model and the most capable in the current lineup. It is designed to think longer per task before responding, which makes it well-suited for complex reasoning, architecture decisions, and long autonomous runs. At approximately $10 per million input tokens and $50 per million output tokens, it is significantly more expensive than other tiers, so pointing it at mechanical or routine work is the fastest way to burn through a usage plan.

What are Claude subagents and how do they save usage?

Claude subagents are cheaper models that execute specific tasks under the direction of a main planning model. Instead of running every step of a complex workflow on the top-tier model, you set one smart model as the commander to plan and make decisions, then delegate execution tasks to Sonnet or Haiku subagents. This matches Anthropic's own developer guidance and can reduce per-task token costs by 2x to 10x depending on how much of your workload is mechanical execution rather than real judgment work.

Got a bottleneck eating your week?

15-minute Resolution Call. I tell you straight if AI can fix it. No pitch deck. No fluff. Live in 7 days from kickoff.

aaarhontoulis@gmail.com  ·  (484) 602-6390