GPT‑5.6 Sol vs Luna vs Terra Pricing Calculator

OpenAI’s GPT‑5.6 family includes three models: Sol, Terra and Luna. GPT‑5.6 introduces three model tiers, multiple reasoning-effort settings and two processing speeds. These options can make it difficult to predict which configuration will deliver the required result at the lowest total cost.

The most expensive model might not be the best and the “cheapest” model might end up being more expensive because of retries etc. 

This article will help to explain the calculator above. 

Limits of the Calculator:

• Current as of July 19, 2026
• There are multiple instances where a more intelligent/expensive model can be lower cost due to finding a more efficient path at completing the task. 
• On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol scored 80 versus Claude Fable 5’s 77.2. OpenAI reports that Sol used less than half the output tokens on this benchmark. A separate Composio benchmark also found lower output-token usage for Sol on its agentic tasks. This suggests greater token efficiency on these specific agentic evaluations, but results may differ across tasks, reasoning levels and agent harnesses.

GPT‑5.6 API pricing

‍

One naming detail matters for developers: the API model name gpt-5.6 routes to Sol. It does not automatically select the cheapest model for each request. Terra and Luna must be selected explicitly with gpt-5.6-terra or gpt-5.6-luna.

‍

These estimates exclude tool-call fees, cache writes and long-context surcharges.

Reasoning tokens can be the largest hidden cost

The output price does not apply only to the answer users can see. OpenAI models may generate hidden reasoning tokens before or between visible outputs, and those reasoning tokens are billed at the model’s output-token rate.

Long prompts cost more

When a GPT‑5.6 request contains more than 272,000 input tokens, OpenAI charges twice the normal input price and 1.5 times the normal output price for the entire request.

This matters for long coding sessions or conversation histories. 

Caching can reduce repeated-context costs

Cached reads receive a 90% discount. However, GPT‑5.6 cache writes cost 1.25 times the normal uncached input rate.

Caching is most valuable when the same large prompt prefix such as documentation will be reused.

Batch and Flex processing offer discounted rates for eligible workloads. See OpenAI’s API pricing for availability and current rates.

Which 5.6 Model and Effort to Use

In our own internal use we have found…

Luna - Luna at high or extra-high effort is a practical daily model but with checks needed for anything complicated.

Terra - Seems to be a reliable model that behaves as you would expect with Medium or High effort. 

Sol - Especially when High or Ultra effort are selected can significantly overthink a prompt and solve the task in a more complex way than necessary. 

Model selection is how “smart” the model is.

Effort is how “hard” the model will work.

Model choice and effort explained like hiring a plumber

The model represents the plumber’s level of expertise and the scale of work they can handle. 

The effort setting controls how thoroughly they plan and investigate before starting and analyze after completing.

• Luna: Best for a clear, routine job, such as clearing a blocked sink.
• Terra: Best for a larger project that requires planning and judgment, such as designing the plumbing for a new house.
• Sol: Best for complex, high-stakes work, such as planning the water and wastewater systems for an entire city.

Effort determines how deeply the model examines the job:

• Light: Check the obvious requirements and use a standard solution.
• Medium: Consider several options and test the most likely approach.
• High: Examine the wider system, compare alternatives and look for hidden problems.
• Extra high: Complete a full analysis before recommending or making changes.
• Max: Spend additional time investigating alternatives, checking results and revising the approach. Max does not have a published fixed token or price multiplier.
• ‍Ultra: Coordinate multiple agents in parallel for demanding work. Ultra can reduce completion time, but typically uses more tokens because several agents may work on the task simultaneously. See OpenAI’s GPT‑5.6 documentation for its explanation of Max and Ultra.

A stronger model can handle a bigger job. Higher effort tells that model to think more carefully about how the job should be done and analyze their work after they completed it.

You would not pay a specialist to fix a loose faucet. But you also would not want someone guessing when water is leaking through your ceiling.

Use the cheapest model and effort level likely to do the job correctly. Choose a stronger model or higher effort when a wrong answer would cost more to fix.

Use Originality.ai’s free Sol vs Luna vs Terra calculator to compare estimated API costs using your own token volume, request frequency and caching assumptions.