Skip to content

What Is LLM Reasoning Effort and How Does It Affect Cost?

TL;DR: LLM reasoning effort settings (low, medium, high) affect how much compute a model allocates to internal reasoning, increasing total token consumption and cost, even when per-token pricing stays the same.

What Is LLM Reasoning Effort?#

  • Many modern AI models offer adjustable reasoning settings such as:

    • Low effort
    • Medium effort
    • High effort
    • Extended thinking
    • Deep reasoning mode

  • These settings control how much internal computation the model performs before generating a response

  • The model uses extra compute to explore multiple reasoning paths, self-check its logic, and refine its answer before outputting anything

How Reasoning Effort Works#

  • When you set a higher effort level, the model spends more time on internal reasoning:

    • Generating internal "thinking tokens" that are not visible in the final response
    • Exploring alternative approaches to the problem
    • Validating its own conclusions before committing to an answer
    • Producing longer, more detailed final responses

  • This is similar to how a person might approach a problem:

    • Low effort: Quick first answer, minimal checking
    • High effort: Multiple solution attempts, verification of assumptions, thorough review before answering

  • The key point: the model does more work internally, and that work consumes tokens

The Two Components of Token Consumption#

  • When using higher reasoning effort, token consumption increases in two ways:

    • Thinking tokens: Internal reasoning tokens generated by the model during its deliberation phase

      • These are consumed by the provider's infrastructure
      • Some providers bill for them, others absorb the cost
      • Example: Claude Sonnet with extended thinking can generate thousands of internal reasoning tokens per query

    • Output tokens: The final visible response tokens

      • Higher effort often produces longer, more detailed answers
      • More examples, explanations, and caveats
      • These are always billed at the output token rate

Cost Impact by Provider#

  • The impact of higher reasoning effort depends on how each provider handles billing

Scenario 1: Token-Based Billing (Most Common)#

  • The per-token price stays the same regardless of effort level
  • However, total cost increases because the model uses more tokens
Effort Level Typical Thinking Tokens Relative Total Cost
Low Minimal (0-100) 1x (baseline)
Medium Moderate (100-1,000) 1.2-2x
High Significant (1,000+) 2-5x+
  • Example: A Claude query that normally costs $0.01 at low effort might cost \(0.03-\)0.05 at high effort due to extended thinking tokens
  • The multiplier varies greatly by task complexity

Scenario 2: Tiered Pricing for Reasoning Modes#

  • Some providers charge different per-token rates for different reasoning modes

  • In this case, two factors compound:

    • More tokens used due to internal reasoning
    • Higher price per token for the reasoning mode

  • Example: Some providers charge a premium for "deep reasoning" or "expert" mode queries

    • Input tokens may be 2x the base rate
    • Output tokens may be 3x the base rate
    • Combined with increased token usage, total cost can be 5-10x baseline

Scenario 3: Fixed Price Tiers#

  • Some platforms offer fixed monthly pricing for different effort tiers
  • E.g., "Standard" vs "Premium Reasoning" subscriptions
  • In this model, the cost is predictable but you pay for the highest tier you need

Concrete Examples#

Example: Mathematical Problem Solving#

  • Task: Solve a complex calculus problem
  • Low effort: Quick answer, may miss edge cases, ~200 output tokens
    • Cost: ~$0.002
  • Medium effort: Shows work, checks key steps, ~500 output tokens + 300 thinking tokens
    • Cost: ~$0.005
  • High effort: Multiple solution approaches, thorough verification, ~1,000 output tokens + 2,000 thinking tokens
    • Cost: ~$0.02

Example: Code Generation#

  • Task: Write a production-grade API endpoint
  • Low effort: Basic implementation, minimal error handling, ~150 output tokens
    • Cost: ~$0.002
  • Medium effort: Includes error handling, input validation, comments, ~400 output tokens + 200 thinking tokens
    • Cost: ~$0.004
  • High effort: Thorough solution with tests, edge case handling, documentation, ~800 output tokens + 1,500 thinking tokens
    • Cost: ~$0.015

Choosing the Right Effort Level#

  • The choice depends on the task and the quality requirements (see How to Compare and Choose LLM Models):

    Low effort is ideal for: - Simple factual queries - Routine code completion - Translation tasks - Quick drafts and brainstorming - High-volume, low-stakes operations

    Medium effort works well for: - Most day-to-day tasks - Code reviews and debugging - Content writing and editing - Data analysis and interpretation

    High effort is best for: - Complex problem-solving - Mathematical proofs and rigorous analysis - Legal or financial analysis - Code that will go to production - Any task where being wrong is costly

Practical Tips#

  • Start low, escalate as needed: Begin with low or medium effort and increase only when the task demands it
  • Match effort to task difficulty: A simple summarization does not need high reasoning effort
  • Monitor your costs: Track token usage across effort levels to understand the real impact
  • Use per-query control: Many providers let you set effort level per request, not just globally
  • Think of it as a dial, not a switch: Experiment with intermediate settings to find the sweet spot for your use case

Conclusions#

  • The relationship between reasoning effort and cost is not always transparent, but it is true that "more reasoning = more tokens = higher cost".

  • I've seen cost multipliers from 1.2x to 10x depending on the provider's pricing model. I match the effort level to the task and check how my provider bills for reasoning before assuming the cost is the same.