GPT-5.6 Luna, Terra, Sol: Which Model to Use & When

OpenAI released the GPT-5.6 family on July 9, 2026, and the structure is different from anything they've shipped before. Not one model. Three, tiered by cost and capability: Luna at the bottom, Terra in the middle, Sol at the top. If you're currently running everything through GPT-4o or a flagship Claude model, you're almost certainly overpaying for a large chunk of your workload.

The opportunity here isn't just "save money." It's routing the right task to the right model so your expensive compute goes where it actually earns its cost. This guide gives you a concrete 3-step plan to do that: audit what you're spending today, test Luna and Terra on the work that doesn't need a sledgehammer, then deploy Sol only where it genuinely pulls ahead.

No fluff. Copy-paste prompts throughout.

Quick Model Overview: Luna, Terra, Sol

Before you audit anything, you need a working mental model of what each tier actually does.

Luna: High-volume, low-complexity tasks

Luna is built for throughput. Fast output, low token cost, no deep reasoning required. Its sweet spot is anything repetitive and well-defined: social captions, product descriptions, email subject line variations, basic summarization, data extraction from structured text. If a task has a clear right answer and doesn't require the model to hold a lot of context or reason across multiple steps, Luna is probably the right call.

Terra: The capable middle tier

Terra sits between Luna's speed-and-cost profile and Sol's frontier reasoning. It's worth testing on work that has more context, requires a consistent voice, or involves moderate complexity — creative writing with specific tone requirements, blog introductions, structured analysis, customer support responses, and code tasks that go beyond simple one-liners. It's more expensive than Luna, but still meaningfully cheaper than Sol. For many teams, a solid portion of the work currently running on flagship models turns out to be well within Terra's range once you actually test it.

Sol: Reserve it for the hard stuff

Sol is for multi-step workflows that require sustained reasoning: complex debugging, automated support flows that branch based on order history, multi-document analysis, anything where the model needs to plan before it executes. It competes with top-tier frontier models. Which also means it costs accordingly. The mistake most people will make is using Sol by default because it's "the best." That's how you burn your budget on tasks Luna could have handled for a fraction of the price.

A quick before/after on cost

Here's the same task run on two different models:

On GPT-4o:

Write a short product description for noise-cancelling headphones.

On Luna:

Create a concise product description highlighting the key features of noise-cancelling headphones.

The output quality for a task this simple is comparable. The cost difference is not. At scale, routing simple content tasks to Luna instead of a flagship model can meaningfully reduce your per-request spend. Note: exact pricing varies by tier and usage volume, so check OpenAI's current API pricing for your workload.

The point isn't one prompt. It's 10,000 prompts. That's where model routing pays off.

Step 1: Audit Your Current Token Usage

You can't route intelligently until you know where your tokens are going. Most people skip this step and just start experimenting, which means they save money in one place and waste it in three others.

Token usage is almost never distributed evenly. In most workflows, a small number of prompt types consume the majority of tokens. Long system prompts, verbose few-shot examples, outputs that include unnecessary preamble, complex summarization tasks run on full documents instead of chunked sections. Find those and you've found your optimization targets.

The audit prompt

If you have access to your API logs or a usage dashboard, run this against your recent prompt history:

Analyze the following list of prompts [PASTE LOGS HERE].

Identify the top 10 most token-intensive prompts. For each one, provide:
1. Approximate token count (input + output combined)
2. A one-sentence description of what the prompt is trying to accomplish
3. Whether the task requires multi-step reasoning, or could be handled by a simpler model

Sort results from highest to lowest token cost.

This gives you a prioritized list of candidates for model switching. The prompts that score high on token count AND low on reasoning complexity are your first migration targets for Luna or Terra.

What to look for

• Prompts with long system instructions that repeat the same context every call (consolidate or cache)
• Output tasks where you're asking for more length than you actually use downstream
• Summarization tasks running on full documents that could be chunked first
• Any task you're currently running on Sol or GPT-4o that fits Luna's profile above

If you want to go deeper on token efficiency before you start switching models, the cache-aware prompting guide covers how prompt caching interacts with token costs in a way most teams overlook.

Step 2: Test Luna & Terra on Routine Tasks

Once you have your audit list, the next step is simple: take the top candidates and run them through Luna and Terra. Compare output quality to what you're getting today. You're looking for "good enough," not "marginally better than the expensive model."

For most high-volume, low-complexity tasks, Luna passes the quality bar. For tasks with more context or moderate nuance, Terra is worth testing before you reach for Sol.

Luna prompts to try

These are the task types where Luna consistently earns its keep:

Write three different Instagram captions promoting our new summer sale.
Keep each caption under 100 characters. Focus on urgency and savings.

Summarize the following customer review in one sentence, identifying
the main complaint or compliment: [PASTE REVIEW]

Extract the following fields from this job posting: job title,
required experience (years), top 3 required skills, salary range if listed.
Return as JSON. [PASTE JOB POSTING]

Run these. Check the output. If Luna's output requires minimal editing, move it off your flagship model permanently.

Terra prompts to try

Terra is worth testing on work with more context, a required voice, or moderate complexity. These are good starting points:

Craft an engaging introductory paragraph for a blog post about the
benefits of remote work. Target audience: young professionals aged 25-35
who are considering negotiating a remote arrangement with their employer.
Tone: conversational, optimistic, grounded in real trade-offs.

Write a Python function that takes a list of dictionaries with keys
"name", "score", and "date", filters out entries with scores below 70,
and returns the results sorted by date descending. Include docstring
and type hints.

A customer just emailed to say their order hasn't arrived and it's been
12 days. Our standard shipping window is 7-10 business days.
Write a response that acknowledges the delay, expresses genuine empathy,
and explains the next two steps they can take (track their order, contact
our team if it doesn't arrive by [DATE]).

If you're consistently getting what you need from Terra on tasks like these, there's no reason to pay for Sol on work in this range. Test before you assume you need the top tier.

If you're building out customer support workflows specifically, the AI agents guide has a useful framework for structuring multi-turn interactions before you decide which model tier to put behind each node.

Step 3: Reserve Sol for Complex Multi-Step Workflows

Sol earns its cost in workflows where the model needs to plan, reason across several pieces of information, and make judgment calls that compound on each other. Not one-shot tasks. Chains.

The most common use case: automated customer support at the complex end of the ticket spectrum. A customer reporting a billing discrepancy tied to a failed subscription renewal that partially charged requires context retention, policy knowledge, and the ability to offer a resolution that accounts for multiple variables. Luna would struggle. Terra might get partway there. Sol handles it.

A tiered workflow example

Instead of routing an entire customer support interaction to Sol, split it by stage:

• Luna: Initial greeting, route the ticket to the right category, extract order number and issue type from the customer message
• Terra: Gather clarifying information, draft the empathetic acknowledgment, pull together what you know about the order
• Sol: Diagnose the root issue, reason across policy + order history + prior interactions, generate the final resolution with three options

Most of the conversation runs on cheaper models. Sol only activates when the complexity actually demands it.

The Sol prompt for this scenario

You are a senior customer support agent for [COMPANY NAME].

Context:
- Customer order #12345 placed on [DATE]
- Item received: [WRONG ITEM NAME]
- Item ordered: [CORRECT ITEM NAME]
- Customer tier: [STANDARD / VIP]
- Prior contacts on this order: [NUMBER]

Task:
1. Acknowledge the error directly and apologize without deflecting.
2. Ask one clarifying question if needed to confirm the resolution path.
3. Offer exactly three resolution options: full refund, replacement shipment
   with priority delivery, or store credit at 110% of order value.
4. Recommend which option is best given the context, with a one-sentence
   rationale.

Tone: warm, direct, no corporate filler.

The structure matters. Sol performs better when you give it a clear sequence of tasks rather than a vague instruction like "handle this customer issue." Break the workflow into numbered steps, specify the output format, and define the tone explicitly.

Other Sol use cases worth testing

• Multi-document analysis where conclusions depend on synthesizing sources that sometimes contradict each other
• Complex code debugging where the bug involves interactions across multiple functions or files
• Data analysis workflows requiring several rounds of reasoning before reaching a recommendation
• Legal or policy document review where edge cases matter and nuance has real consequences

For data-heavy workflows where AI is doing the heavy scanning and you're making the final call, this piece on AI processing ops data while keeping strategy decisions yours is worth reading alongside your Sol experiments.

Frequently Asked Questions

How much cheaper is GPT-5.6 Luna compared with GPT-4o?

OpenAI hasn't published a fixed percentage comparison, and pricing varies by input vs. output tokens and usage tier. The practical guidance: treat Luna as your high-volume, cost-efficient tier and benchmark your specific prompt types against GPT-4o to find your actual savings. Don't rely on a number you read here, including any estimate from the original announcement. Check the OpenAI API pricing page for current rates.

When should I use Terra instead of Luna?

Use Terra when your task involves more context, tone-matching that goes beyond basic instructions, moderate reasoning, or creative output that needs to hold a specific voice. If Luna's output consistently needs heavy editing, that's a signal to move to Terra. If Terra's output also needs heavy editing, test Sol, but first check whether your prompt itself needs work before upgrading the model.

Can Sol hold its own against other top-tier frontier models on complex tasks?

Sol is positioned to compete at the frontier level, and benchmarks show it performs alongside other leading models on demanding tasks. Whether it outperforms, matches, or falls short on your specific work depends on the task. Benchmarks give you a starting point, not a final answer. For high-stakes workflows, run your actual prompts on both models, score the outputs against your real criteria, and let that data drive the decision. The Claude Sonnet prompt testing guide has a solid framework for running that kind of head-to-head comparison.

What's the best way to test token usage before switching models?

Pull your API logs from the past 30 days if you have them. If you don't have log access, run your 10 most common prompt types through a token counter (OpenAI's tokenizer playground works) and tally input plus expected output. That gives you a usage map. Then run the same prompts through Luna and Terra and compare output quality side by side before committing to a switch.

Are GPT-5.6 Luna, Terra, and Sol available through the OpenAI API?

Simon Willison's coverage from July 9, 2026 confirmed the release of the GPT-5.6 family. For current API availability, model IDs, and access tiers, check OpenAI's API documentation directly, as availability can vary by account type and region.

The Model Isn't the Bottleneck. The Routing Is.

The GPT-5.6 family doesn't change what good prompting looks like. It changes where you send the prompt. Most AI costs aren't high because the models are expensive. They're high because people route everything to the most capable model out of habit, then wonder why their API bill keeps climbing.

Audit first. Test Luna and Terra on the work in your stack that doesn't need frontier reasoning. Save Sol for the cases where the stakes and complexity actually warrant it. That's the whole playbook.

If you want structured prompt templates already optimized for different task complexities and use cases, Ultra Prompt has a library built for exactly this kind of work.