GPT-5.6 Sol, Terra, and Luna: What’s the Difference?

The new GPT models are out and about, and people are understandably confused about what each version actually means.

At least I am.

We’ve gone from GPT-5 to GPT-5.5, and now GPT-5.6 has shown up with three new names: Sol, Terra, and Luna. Are these different versions? Different personalities? Different models for different jobs?

Basically: yes.

The easiest way to think about GPT-5.6 is as a three-model lineup:

• Sol is the powerful one.
• Terra is the balanced, everyday one.
• Luna is the fast and affordable one.

They are all part of the GPT-5.6 family, but they are designed for different tradeoffs between intelligence, speed, and cost.

The quick version

Those are OpenAI’s preview prices, so they could change later. OpenAI’s current pricing and access information

So what does “GPT-5.6” mean?

The 5.6 is the generation. The names are the tiers.

OpenAI says Sol, Terra, and Luna are meant to be durable capability levels that can improve on their own schedules. In normal-person language, that probably means we shouldn’t assume Sol, Terra, and Luna will always be released as one perfectly synchronized group.

Instead, OpenAI seems to be building a lineup where you pick the model based on what you are trying to do.

Need the smartest available option? Use Sol.

Need something strong but reasonably priced? Use Terra.

Need to answer a lot of simple requests quickly? Use Luna.

It is a little like choosing between a race car, a reliable sedan, and a fuel-efficient compact. They are all cars, but you probably would not use the race car to drive to the grocery store.

GPT-5.6 Sol: the one you call when things get complicated

Sol is the flagship model.

It is designed for the jobs where the model has to think through several steps, use tools, deal with mistakes, and keep working toward a larger goal.

OpenAI is highlighting Sol for:

• Complex software engineering
• Long-running agent tasks
• Scientific research
• Biology and genomics work
• Advanced cybersecurity analysis
• Computer-use workflows

Sol also introduces a new reasoning setting called max, which gives it more time to work through difficult problems. There is also an ultra mode that can use subagents to divide up complicated work.

That is the part that makes Sol feel less like a chatbot and more like a junior technical team you can point at a problem.

You might give it a large coding task and ask it to inspect a repository, plan the changes, test the result, and fix anything that breaks. Or you could use it for a research problem where the first answer is not going to be good enough.

OpenAI says Sol improves on GPT-5.5 in coding and biology benchmarks, and describes it as its most capable model yet for cybersecurity. OpenAI’s GPT-5.6 Sol announcement

The downside is that Sol costs more and will probably take longer on difficult tasks. It is powerful, but you would not want to use it for every tiny request.

Sol is probably the right choice when:

• The task is genuinely difficult.
• You need multi-step reasoning.
• You are working with code, tools, or files.
• A bad answer could create expensive problems.
• You would rather wait longer for a better result.

Sol is the “okay, bring in the expert” model.

GPT-5.6 Terra: probably the model most people will actually use

Terra is the middle option, and it may end up being the most useful one for everyday work.

OpenAI describes Terra as a balanced model for normal professional tasks. It is intended to offer performance competitive with GPT-5.5 while costing about half as much.

That puts Terra in a pretty attractive position.

It should be capable enough for:

• Coding help
• Business writing
• Research
• Planning
• Document analysis
• Customer-support workflows
• General-purpose AI applications

Terra is not necessarily the “cheap version” in the way people sometimes think of a mini model. It is more like the model you use when you want strong results without paying flagship prices for every request.

For many businesses, Terra will probably be the default. Sol can handle the complicated cases, while Terra takes care of the normal workload.

Terra is probably the right choice when:

• You need a strong general-purpose model.
• You are building a production application.
• You care about both quality and cost.
• Your tasks are important but not extremely specialized.
• You want something more capable than a lightweight model.

Terra is the sensible daily driver.

GPT-5.6 Luna: the speedy one

Luna is designed to be fast and inexpensive.

It is the model you use when you need to handle a large number of requests without spending a fortune.

That could include:

• Summarizing text
• Rewriting content
• Classifying support tickets
• Extracting information
• Answering simple questions
• Generating short responses
• Powering high-volume chatbots

Luna costs much less than Sol, which matters when an application is handling thousands or millions of requests.

It would be wasteful to use the most expensive model in the family every time someone asks for a paragraph to be shortened or an email to be rewritten. Luna is meant for that kind of work.

That does not mean Luna is useless or unintelligent. It is still part of the GPT-5.6 family. It is just optimized for speed and cost rather than maximum reasoning depth.

Luna is probably the right choice when:

• You need quick responses.
• You are processing a lot of requests.
• The task is relatively straightforward.
• You can retry or escalate difficult requests.
• Keeping API costs under control matters.

Luna is the “get it done quickly” model.

The smart way to use all three

The interesting part is that you do not necessarily have to choose only one.

A real application could use all three models:

Simple request
      ↓
GPT-5.6 Luna

Normal business task
      ↓
GPT-5.6 Terra

Complicated or high-risk task
      ↓
GPT-5.6 Sol

For example, a customer-support system could use Luna to sort incoming tickets, Terra to draft ordinary responses, and Sol to handle unusual or sensitive cases.

That is probably where this model lineup makes the most sense. Instead of asking, “Which GPT model is the best?” the better question becomes, “Which model is best for this particular job?”

What about the new caching features?

GPT-5.6 also brings changes for developers who repeatedly send the same large prompts.

OpenAI says the family supports more predictable prompt caching, including explicit cache breakpoints and a minimum cache life of 30 minutes. Cached input reads still receive a 90% discount, while cache writes cost 1.25 times the normal uncached input rate.

That could be useful for applications that repeatedly include the same:

• Company knowledge base
• System instructions
• Tool definitions
• Policy documents
• Product documentation

If you are building an AI assistant that always needs access to the same large set of reference material, caching could make those repeated requests cheaper.

The cybersecurity part is a big deal

GPT-5.6 is also getting attention because of its cybersecurity capabilities.

OpenAI says the models are better at finding and fixing vulnerabilities, while safeguards are intended to make dangerous end-to-end attacks harder to perform.

The company has added multiple layers of protection, including model training, real-time monitoring, account-level signals, access controls, and automated red-teaming.

Sol, Terra, and Luna are all classified as having high cybersecurity capability, although OpenAI says they remain below its “critical” threshold. In its testing, Sol found bugs and pieces of potential exploits but did not autonomously produce a complete working exploit under the tested conditions. GPT-5.6 Preview System Card

Of course, benchmark tests are not the same thing as a guarantee. A model can be combined with tools, scripts, databases, and human expertise. That is why the release is being handled carefully.