Four Ways to Deal with Claude's Reduced Usage Limits

Claude’s usage limits got nerfed hard a month ago.

The nerf was so hard that I had to switch from Opus to Sonnet as my default agent (and lose the flavour of a persona that I created and liked). So I’m genuinely upset about it.

But I have to get over it and move on.

Along the way, I discovered a few keys that helped me continue to use Claude even with the reduced limits.

They are:

1. Default to Sonnet
2. Reduce Skills and MCP Tool usage
3. Upgrade to Max for better cache TTL (if you want)
4. /clear and /compact aggressively
5. Use Codex as a supplement

Default to Sonnet

Opus is way better but with the new usage limits, there are only so many things you can do with Opus before you run out.

So defaulting to Sonnet is nothing but a necessity.

But in my bid to make Sonnet work, I discovered that Sonnet is great when you don’t need it to help you think hard about things.

But these tasks are better left for Opus:

1. Thinking through problems
2. Debugging hard problems
3. Creating detailed plans
4. Drafting words with a flavor

So balancing between Sonnet and Opus is a trade-off decision, and I had to learn when Sonnet was enough.

What’s interesting also is that:

• Low-effort Opus is generally not worth it at all
• Medium effort Opus is generally okay for most of the harder tasks I would like it to do
• High-effort Sonnet is still worse off than low-effort Opus
• I am usually on medium effort Sonnet

Some people say you should use Haiku as your main agent because it can “delegate” to smarter ones… I honestly do not recommend it — I vomit blood every single time I try talking about the haiku.

Startup Costs

It’s worth noting that Opus, Sonnet, and Haiku have different startup costs that depends on their own system prompts.

Here are /context screenshots to prove the values I used above are accurate as of 1st May, 2026

While taking these screenshots, I also noticed a few interesting things:

1. Haiku system tools are 20.6k tokens! (Whoa!)
2. The cost for custom agents and messages is higher in Opus compared to Sonnet and Haiku — even when their values are completely the same!

What this also means is that agents definitions costs more on Opus!

That leads nicely to my next point.

Reduce Skills, Agents, and MCP Tool definitions

Agents, skills, and MCP tool definitions cost context tokens. They are charged for every single conversation. So you’re always paying for them even if you don’t use them at all.

Before I knew this, there was a time where:

• My MCP tools went up to 30,000 tokens
• My skills up to 5,000 tokens

Which means I’m paying 35,000 tokens extra for any conversation, which costs even more each turn…

The best action here is not to eliminate the use of skill and MCP tool definitions altogether because those are the very things that make Claude versatile.

But you want to consider which tools are necessary and which ones are not.

• Eliminate the ones you don’t use
• Keep the ones you always use

For those that are in between, it’s possible to reduce token usage with a skill router and an agent router. I’ll talk about that in a future article.

For MCP it’s slightly easier. Opus and Sonnet come with an option to lazy load MCP tools so you can just enable it with this setting. (Just ask Sonnet to help you do it).

{
  "env": {
    "ENABLE_TOOL_SEARCH": "true"
  }
}

Haiku doesn’t support the lazy loading of MCP tools. That’s why MCP costs are added to Haiku.

Upgrade to Max for better cache TTL (if you want)

Caching is very important when it comes to saving cost when using LLMs, because the cost for reading a cache is 0.1x the usual cost.

Anthropic has 2 different cache mechanisms:

• 5 minutes TTL (writing cache costs 1.25x)
• 1 hour TTL (writing cache costs 2x)

Unfortunately, Pro subscriptions (which I use) is limited only to the 5-minute cache.

So I don’t multitask with many agents because there’s a high chance of missing the hash.

If you are someone who prefers to fire up many agents and multi-task between them, then I highly recommend upgrading to Max because that is the only way you get a one-hour cache.

Otherwise it’s best to change how you work with LLMs so you can stay within the chat times.

Clear and Compact aggressively

Many people have made a lot of noise about this so I’ll not add to the noise.

Instead, here are things that people have never said before:

1. When you miss a cache, I think it’s a good time to compact or clear.
2. This is a hypothesis: I think ing within the 5mins cache allows you to start your next session at the cached rate, but the downside is you lose the history.

Use Codex as a supplement

I know some people will roll their eyes at this recommendation but I honestly found it useful.

That’s because when Claude gets stuck on a task, it remains stuck on a task anyway, so you’re burning tokens while it’s going around in circles.

In these cases it is best to use a different set of eyes. And the best two eyes that I can think of right now are:

1. Use Codex with high or x-high.
2. Roll up your sleeve and use your own eyes

That’s it! Hope you found this useful!