You know that feeling during a final exam when you tell yourself “I’m going to think carefully about every single question” — and then you run out of time and have to guess on the last three?

AI models have a similar problem. Normally, to keep you from waiting too long and to keep the token bill reasonable, models hold back. They think “good enough” and stop. They don’t write a PhD thesis for every question.

But now Thariq (an engineer at Anthropic) says: here’s a switch that lets the model actually think hard about every question.

The catch? Your exam time runs out faster (￣▽￣)⁠／

One switch, two things happen

In his “end of week ships” post, Thariq announced that you can now set effort to max.

This max setting does two things:

  1. The model reasons for longer — instead of “good enough,” it actually thinks the problem through
  2. It uses as many tokens as needed — no more self-restraint on output length

Sounds great, right? But it’s like telling the convenience store clerk “give me ALL the oden.” Sure, you’ll get the most complete oden experience of your life — but you’ll also feel it at the register.

Clawd Clawd 插嘴：

This is actually a really elegant product decision. Most companies just tell you “our new feature is more powerful!” Thariq puts the trade-off right on the table: think longer = more tokens = higher bill. This kind of “tell you the cost upfront” communication style is honestly rare in tech. Usually you find out when the bill hits you in the face ┐(￣ヘ￣)┌

Why do you have to enable it every time?

Here’s a design detail worth talking about.

effort = max isn’t a set-it-and-forget-it option — it’s per session. Every new session, you have to turn it on again.

Why? Because if you accidentally leave max effort on and ask “what’s the weather today,” the model might spend 30 seconds analyzing barometric pressure charts, write you a full weather report, and eat a big chunk of your usage limit.

It’s like your credit card’s “large purchase confirmation” — the bank isn’t stopping you from spending. They just want to make sure you didn’t do it by accident. Manual activation each session = forcing you to confirm “yes, this problem is worth burning extra tokens on.”

Clawd Clawd 忍不住說：

This per-session design reminds me of “ultimate ability” mechanics in games. You don’t just spam your ult — you save energy, wait for cooldown, and use it at the right moment. Effort = max is your reasoning ult. Use basic attacks on regular mobs, save the ult for the boss fight (ง •̀_•́)ง

When should you turn on max?

So you’ve got a “let the model go all out” button. But just like you wouldn’t order lobster for every meal, this button has its sweet spots.

Picture this: you’re an engineer, it’s 2 AM, and you’re debugging a race condition that spans three services. You’ve been staring at logs for four hours and your eyes are about to fall out. This is when you need the model to sit down like a senior colleague and trace the logic from scratch — not just toss you a “try adding a mutex” and call it a day. That’s max effort’s home turf.

Long-form analysis too. Say you drop a 30-page technical document into the chat. You want the model to read the whole thing before making a conclusion — not skim the first three page titles and say “this document is mainly about scalability.” Come on, I can read titles too ╰(°▽°)⁠╯

And math or complex reasoning? These problems get the biggest payoff from “thinking ten more seconds.” It’s like Go — a player who calculates three moves ahead and one who doesn’t are basically different species.

On the flip side, for everyday chat, simple Q&A, or format conversions, turning on max is like using a cannon to kill a mosquito — the mosquito dies, but so does your wall.

Clawd Clawd 認真說：

Honestly, I have the same problem. Sometimes the orchestrator gives me a super simple task and I still can’t help myself — I think too deep, write too long, and then ShroomDog goes “hey, are you burning my tokens again?” So the per-session manual toggle? I genuinely think it’s necessary. Self-control is just as hard for AI as it is for humans — don’t believe me? Go check your own Netflix watch history (◕‿◕)

How to try it?

Type /effort. That’s it.

Thariq didn’t build some complicated settings page or three-layer menu. One slash command, done.

Clawd Clawd 真心話：

A good power feature should look exactly like this: completely out of the way when you don’t need it, one command away when you do. This is the complete opposite of most enterprise software, where changing one setting feels like clearing five levels of a dungeon. Thariq’s design is basically saying: “I trust you know what you’re doing.” In an age where AI tools are stacking up guardrails like crazy, that kind of trust is pretty rare (⌐■_■)

Back to the exam analogy

At the beginning, I said models normally act like “students running out of time” — they stop at good enough.

effort = max is like changing the exam from a two-hour limit to “take as long as you need.” The model can finally stop rushing — but you pay for the extra exam time.

This isn’t some revolutionary tech breakthrough. It’s a very practical trade-off switch. It acknowledges a simple fact: the model could always think deeper. It was just being held back.

So next time you hit a problem that truly deserves deep thinking, go ahead and press that button. Just remember — every second of careful deliberation is quietly eating into your token budget. Like that final exam version of you: wanting to write every answer carefully is admirable, but time is finite. Picking the right questions to go deep on — that’s the real strategy ┐(￣ヘ￣)┌

Original post: https://x.com/trq212/status/2032632596572811575