You're wasting your AI credits on Max thinking

GPT-6 Astra is smarter, but that does not mean every task needs maximum reasoning.

Since GPT-6 Astra launched, I’ve noticed the same pattern while talking to people about how they use it.

They choose Astra because it is the most intelligent model available. Then they select Max or Ultra thinking because more thinking must produce an even better answer.

That assumption makes perfect sense.

Everywhere else in life, taking more time to think usually improves the result. We tell people to slow down, consider their options and avoid rushing important decisions. When an AI product offers effort levels from Low through to Max or Ultra, the labels naturally feel like a scale running from worse to better.

But AI reasoning effort does not work that neatly.

The model determines the intelligence available. The effort setting determines how much time and compute it can spend applying that intelligence.

Astra on Low is still Astra. Max does not make it a smarter version of itself. It gives the same model more room to plan, explore alternatives and check its work.

Sometimes that is valuable. Often, it is an expensive way to complete a simple task.

Key takeaways

• Model choice and reasoning effort are separate decisions.
• Higher effort creates an opportunity for better reasoning. It does not guarantee it.
• Clear requirements and useful context matter more than a prestige setting.
• Use the lowest effort likely to complete the task reliably.
• Increase effort when complexity, uncertainty or the cost of failure justifies it.

1. What does reasoning effort actually change?

Reasoning effort guides how much internal work a model performs before and during its response.

At lower effort, the model prioritises speed and reduced token use. At higher effort, it has more room to break down the problem, inspect alternatives, use tools and reconsider its approach.

OpenAI’s reasoning documentation describes the trade-off between speed, token use and potentially higher-quality responses. The important word is potentially. Higher effort creates an opportunity for better reasoning. It does not guarantee a better result.

The model also adapts to the task. A simple question may use relatively little reasoning even when a higher setting is selected. A complex task can consume far more. The setting guides the available effort rather than prescribing a fixed amount of thinking.

Those reasoning tokens may be invisible to the user, but they still count towards usage. In the API, they occupy context and are billed as output tokens. Subscription products use their own usage and credit systems, but the practical lesson is the same: additional reasoning is not free.

OpenAI recommends the highest effort levels only when evaluation shows a clear benefit that justifies the added latency and cost.

That is very different from “always choose Max for the best answer”.

2. Intelligence and effort are separate decisions

It helps to think of AI settings as two different controls.

The first control selects the model. This determines the model’s underlying capability: how well it follows instructions, handles ambiguity, works across a long task and reasons about difficult problems.

The second control selects the effort. This determines how much reasoning budget that model can apply to the current task.

These controls interact, but they are not interchangeable.

A less capable model at Max does not automatically become equivalent to a more capable model at Low. Likewise, choosing the strongest model does not mean you should automatically give it the largest possible reasoning budget.

This is why “Astra Low” should not sound like a compromised version of Astra. You are still getting Astra’s underlying capabilities. You are asking it to reach an answer with less deliberation.

For many ordinary tasks, that is exactly what you want.

3. Why can more thinking produce a worse result?

More reasoning gives the model additional opportunities to find a good solution. It also gives it additional opportunities to pursue the wrong one.

If your instructions are unclear, Max effort cannot discover requirements you never provided. It may confidently develop one interpretation much further than necessary.

If the task is already simple, additional reasoning may introduce needless complexity. The model can expand a small solution, search for problems that do not exist or second-guess a correct initial answer.

If your request has no clear stopping point, a high-effort model may keep researching, analysing and checking without materially improving the result.

More thinking cannot compensate for:

• Missing context
• Conflicting instructions
• An unclear outcome
• Incorrect source information
• Missing tools or permissions
• No way to verify the answer

Before increasing effort, improve the task. Explain what success looks like, provide the relevant information and tell the model how to check its work.

A clear prompt at Medium will often outperform a confused prompt at Max.

4. When should you use each effort level?

There is no universal setting for every model and task, but a simple ladder covers most situations. Product labels vary: for example, the Astra API supports Low, Medium, High, Extra High and Max, while eligible ChatGPT Work accounts may also offer Ultra.

Low: the task is clear and easy to verify

Use Low for straightforward work with a narrow outcome.

Examples include:

• Rewriting a paragraph
• Summarising supplied text
• Extracting information from a document
• Formatting data
• Drafting a routine message
• Making a small, clearly specified change
• Finding a known fact using an appropriate source

These tasks do not need extensive exploration. You can also inspect the result quickly and request a correction if needed.

Medium: some planning or judgement is required

Medium is the sensible default for normal professional work.

Use it when the model needs to connect several pieces of information, follow multiple steps or make modest judgement calls.

Examples include:

• Comparing several options
• Researching a defined question
• Preparing a structured document
• Analysing a spreadsheet
• Planning a small project
• Implementing a scoped feature
• Reviewing information against clear criteria

Medium gives the model room to reason without treating every request like an unsolved research problem.

High: the problem is genuinely difficult

Use High when the task contains meaningful ambiguity, interacting constraints or expensive failure modes.

Examples include:

• Diagnosing an intermittent technical problem
• Designing an important system or workflow
• Evaluating competing strategies
• Planning a sensitive migration
• Reviewing legal or financial material with appropriate human oversight
• Investigating a question across many sources
• Finding subtle security or reliability risks

At this level, additional planning and checking can materially change the result.

Extra High or Max: you can explain why High is insufficient

The highest effort settings should have a specific reason behind them.

Use it when the task involves a long chain of dependent decisions, several plausible solutions that need testing, an unusually high cost of missing an edge case or a previous attempt that failed because the model needed more room to reason.

“Because I want the best answer” is not specific enough.

Use Max when you can complete this sentence:

This task needs additional reasoning because…

If you cannot name what the extra reasoning must accomplish, start lower.

5. Ultra is not merely “even smarter”

Ultra deserves special attention because its name strongly implies a superior answer.

In ChatGPT Work, Ultra uses maximum reasoning and can proactively delegate suitable work to subagents. That means the system may divide a request into parts and have several agents work in parallel.

That suits large research projects, complex investigations and work with genuinely independent streams.

It is unnecessary for rewriting a paragraph, summarising one document or making a routine change.

Selecting Ultra for a small task can be the equivalent of assembling a project team to rename a folder. The team may do an excellent job, but that does not make it a sensible use of resources.

6. Use difficulty signals, not prestige

People often select the highest setting because the task feels important.

Importance matters, but it is not the same as reasoning difficulty.

A one-sentence public announcement may be important, yet simple to draft and easy to review. A harmless logic puzzle may be unimportant, yet require substantial reasoning.

Effort should follow:

• Uncertainty: How unclear is the problem?
• Dependencies: How many decisions affect one another?
• Verification difficulty: How hard is it to determine whether the answer is correct?
• Cost of failure: What happens if the model misses something?
• Task length: Must the model remain coherent across a sustained workflow?

If those factors are low, additional effort is unlikely to earn its cost.

7. Increase effort only after diagnosing the problem

When an answer is poor, many users immediately move the effort slider upwards.

That should not be the first response.

First ask why the answer failed:

• Did the model misunderstand the request?
• Was important context missing?
• Did it use the wrong source?
• Was the requested outcome ambiguous?
• Did it need a tool it did not have?
• Did you define how the work should be verified?
• Was the underlying model capable enough for the task?

If the failure came from weak instructions, fix the instructions.

If it came from missing information, provide the information.

If it came from insufficient model capability, change the model.

Increase reasoning effort when the model understood the task but needed more room to work through its complexity.

8. The goal is enough thinking

GPT-6 Astra is an extremely capable model. That is exactly why it should not require Max or Ultra for every interaction.

The mistake is understandable: the product presents effort settings in an order that looks like a quality ladder. Low appears worse than Medium, while Max and Ultra appear to promise the best possible result.

A better way to use them is as a budget.

Start with the lowest effort that is likely to complete the task correctly. Make the requirements clear. Give the model the context it needs. Define how the result should be checked.

Then increase effort when the task demonstrates a genuine need for deeper reasoning.

The goal is enough thinking to finish the job reliably.