How to use Reasoning Effort?

What is Reasoning Effort?

Reasoning Effort controls how much “thinking” a model does before it gives you an answer. You can set it to low, medium, or high.

• Low → Quick answers, but less detailed reasoning
• Medium → Balanced speed and depth (default)
• High → Deep reasoning and more complete answers, but slower and more expensive

This parameter is especially useful for complex problem-solving, coding tasks, or math/logic problems.

How does Reasoning Effort work?

When a model answers, it generates hidden “reasoning tokens” (sometimes called chain-of-thought).

• Without adjustment → The model uses a default amount of reasoning.
• With Reasoning Effort → You decide how much effort it should spend “thinking” before finalizing the answer.

That means:

• Less effort = faster, cheaper, but shallow responses
• More effort = slower, costlier, but deeper and more accurate

How to use Reasoning Effort with OpenAI models

When calling an OpenAI reasoning model (o1, o3, o3-mini, GPT-5, etc.), you can set:

"reasoning_effort": "low"   // options: low, medium, high

Example:

response = client.chat.completions.create(
    model="o3-mini",
    messages=[{"role": "user", "content": "Solve this math problem: 243 / 9"}],
    reasoning_effort="high"
)

How to use Reasoning Effort with Claude models

Claude 3.7 Sonnet and later also support variable reasoning depth.
You can set the level of “thinking” tokens in a similar way via API (Anthropic docs recommend adjusting based on task complexity).

Key things to know

• Defaults: Medium is default if you don’t set anything.
• Trade-offs: Higher effort improves reasoning tasks (math, coding, logic puzzles), but takes longer and costs more.
• Performance: Benchmarks show accuracy can improve 10–30% at high effort, depending on the model and task.
• Cost: Reasoning models already cost more; “high” effort multiplies this. One study found they can be 10–74× pricier than standard models.

Example use cases

• Low → Fast Q&A, customer support, summarization
• Medium → Most general use cases
• High → Advanced coding, multi-step math, research, reasoning benchmarks

Reasoning Effort Pricing (Conceptual)

Pricing varies by model, but here’s the pattern:

‍

When to use Reasoning Effort?

• When accuracy matters more than speed
• When handling complex reasoning, math, or coding problems
• When testing models against benchmarks or evals
• When you want to experiment with cost/performance trade-offs