Terra and Luna: Fast, Affordable Reasoning
On this page
The Reasoning You’ll Use Every Day
In the previous generation, fast and affordable reasoning was handled by o4-mini. That model left the API on October 23, 2026, with gpt-5.6-terra as its designated successor. In the current lineup, that role falls to two models: GPT-6.1 Terra, which is fast and efficient, and GPT-6 Luna, the most economical option. Paired with moderate effort settings, they handle the bulk of daily technical tasks and deserve a prime spot in your workflow.
Where Terra and Luna Fit In
In the API, GPT-6.1 Terra (gpt-6.1-terra) costs $1 per million input tokens and $5 per million output tokens. OpenAI positions it as a fast, efficient model for visual reasoning, precise computer control, and clear writing. GPT-6 Luna (gpt-6-luna), priced at $0.10 and $0.50, is built for targeted, high-volume repetitive tasks: extraction, classification, transformation, and structured summaries. Both accept all effort levels from none to max and share the rest of the lineup’s 1.05 million token context window.
In ChatGPT, you’ll mostly encounter them in Work and Codex. OpenAI recommends Terra for writing, image understanding, and working within applications, while Luna is best for clean, repetitive tasks where you can easily recognize a good result; for Luna, the documentation suggests starting at the High level. In Chat, the Free and Go plans rely on a Luna model, with the Think button available for harder questions.
GPT-6.1 Sol and GPT-6 Astra retain the advantage for highly abstract problems, strategic trade-offs involving many factors, exhaustive explorations, and long-form work where response time matters less.
When to Turn Up the Dial
The key question isn’t “which model should I choose?” but “what effort level should I activate?” Four signals justify moving from none or low to medium on Terra: the task involves calculation or logical verification; it breaks down into multiple steps; the response at the lowest level was approximate or incorrect; or the conclusion requires cross-referencing multiple pieces of information. If high still isn’t enough, it’s time to switch to GPT-6.1 Sol.
Five Use Cases Where This Setting Shines
As a daily coding assistant, Terra set to medium reasons through program logic, spots bugs, and proposes well-argued fixes: “Here’s a Python function that should calculate the median of a list, but it sometimes returns the wrong result. Find the bug and fix it: [code].” For quick data analysis, it avoids the need to invoke a more expensive model: “Here are sales figures for 5 products over 6 months. Which has the best trajectory, and which is at risk of dropping off? Justify with calculations.”
It also excels as a verifier, whether for your own calculations or another model’s output. Thanks to its visual capabilities, it can comment on a dashboard screenshot: “What are the three most important observations, and what do you recommend?” Just keep in mind that numbers read from an image still need verification. Finally, for repetitive tasks like sorting 500 tickets or extracting fields from 200 documents, Luna set to low or medium offers the best quality-to-cost ratio; test on a sample first.
Three Habits That Change Everything
Be direct: “Calculate the compound annual growth rate for this revenue: 2021 = €1.2M, 2022 = €1.5M, 2023 = €1.9M, 2024 = €2.3M, 2025 = €2.8M” works better than a long preamble. OpenAI recommends simple, direct instructions for its reasoning models. Ask for verification within the result: “Give the result, then verify it using a different method.” And work in short iterations: the speed of these settings is designed for rapid back-and-forth.
Your Turn
Four tasks, one stopwatch. Test Terra at medium (in the API or Work), or the Medium level in Chat as a fallback, on these four requests. Note the response time and verify each result.
- “Solve: if 3x + 7 = 2(x − 4) + 15, what is x? Verify.” (Expected answer: x = 0.)
- “Here is a JavaScript code snippet: [code with a subtle bug]. Find the bug.”
- “My quarterly revenue: Q1 = 120k, Q2 = 135k, Q3 = 128k, Q4 = 142k. Calculate the trend and project the next Q1.”
- “Software costs €15,000 per year. The open-source alternative costs €5,000 to deploy, then €2,000 per year for maintenance. When does the alternative become cost-effective?”
Each is handled quickly with sufficient reasoning. Using a more expensive model at a high effort level for this type of request costs more and takes longer, adding no value.
Official sources: Models (OpenAI API: GPT-6.1 Terra and GPT-6 Luna, effort levels, pricing) · Deprecations (o4-mini removed October 23, 2026, successor) · Models (ChatGPT documentation: Terra and Luna use cases in Work, starting level) · Reasoning best practices (direct instructions)
Key Takeaways
- o4-mini left the API on October 23, 2026; GPT-6.1 Terra and GPT-6 Luna now handle fast reasoning
- Terra: $1 / $5 per million tokens for writing, images, and application control; Luna: $0.10 / $0.50 for high-volume repetitive tasks
- Both accept effort from
nonetomax; increase effort before switching models - Use direct instructions, request verification in the output, and iterate quickly
- Reserve GPT-6.1 Sol and GPT-6 Astra for abstract, strategic, or exhaustive problems