You open a model picker and see , Sol, Terra and Luna. Some entries say ; another says . Then there are settings such as medium and high. Are these separate models, nicknames for one model, or operating modes?
Start with one distinction: Sol, Terra and Luna belong to ; appears as . They are not all . Nor must you choose between calling a model and calling it : both can be parts of its full name. This guide uses official documentation checked on September 8, 2026.
Before the name: what is an AI model?
A model is a trained mathematical system that transforms inputs into outputs. For a language model, the input may contain a question, document or instructions, and the output may be an explanation, summary or code. GPT stands for Generative Pre-trained Transformer, terminology covered in ’s key concepts.
Training adjusts numerical values called parameters or weights. These values influence how the system processes information. They are not a collection of answers individually written by a person. The model learns patterns it uses to produce new responses. A convincing response can still be wrong.
Text is represented using : pieces that may correspond to a word, part of a word or punctuation. The system processes context and generates an output in steps. Running the trained model is called inference. Providing a document changes the context for that answer; by itself, it does not train a new model.
For example, when you ask it to summarize three emails, the model contributes its learned capabilities and the emails supply the information for that task. The answer depends on both. A more impressive model name does not automatically supply a missing fact.
How to read GPT‑5.6 Sol
Read the name in parts: GPT identifies the line, 5.6 identifies its version or family, and Sol distinguishes a model within that family. In , the number is 6 and the name is .
Think of a computer series sold in several configurations. Sharing a series does not imply identical capabilities or prices. This analogy helps you read a catalogue; it does not describe a model’s internal construction.
The number is not an intelligence measurement either. Six does not mean six units of capability, and it does not tell you how much better the model will perform than 5.6 on your task. It is a product designation. Establish concrete differences through documentation and evaluations relevant to your work.
Sol, Terra and Luna within GPT‑5.6
documents three models with different identifiers. Their names distinguish options with different capability and cost priorities within the same family.
Sol: the family’s flagship option
GPT‑5.6 Sol is positioned for complex professional work. roughly compares its role with the unsuffixed model tier in earlier families. Its explicit identifier is gpt-5.6-sol.
One detail explains some of the confusion: the gpt-5.6 alias routes requests to Sol according to that documentation. An alias is another identifier pointing to a model. In this case, it does not mean a request arbitrarily chooses among Sol, Terra and Luna.
Terra: balancing capability and cost
GPT‑5.6 Terra is positioned to balance capability and cost. The documentation roughly compares it with the earlier mini tier. Its identifier is gpt-5.6-terra.
Roughly matters: the comparison describes a product’s position in a range. It does not mean Terra is an old mini model with a different label.
Luna: cost-sensitive work at volume
GPT‑5.6 Luna targets high-volume work where cost is a major consideration. roughly compares its position with the earlier nano tier. Its identifier is gpt-5.6-luna.
The name alone does not reveal a parameter count or the internal changes distinguishing it from Sol or Terra. A position in a product catalogue is not an architectural specification.
Why is GPT‑6 called Astra?
The number and the name are two parts of the same designation. GPT‑6 Astra is already listed under that full name, with the identifier gpt-6-astra. positions it as its most capable model for complex work from start to finish.
does not subsequently transform into . A picker may use an abbreviated label, but check the full name when comparing options or saving a configuration.
These four names do not establish a rule for predicting future products. The existence of , Terra and Luna does not prove there will be a Terra or Luna. That would require an official announcement or model page. The current catalogue describes available models, not a release calendar.
A model, an application and reasoning effort
and are products through which you can use models, with interfaces and tools. The model is one part of that experience. Searching the web, reading a file or running a command also depends on available tools and permissions.
Reasoning effort is another setting. The model-picker documentation distinguishes selecting a model from adjusting its reasoning effort. Higher effort can help with complex tasks while using more time and tokens. Changing Sol from medium to high keeps the model as Sol. It does not turn it into Terra or .
An agent adds a working process: it uses the model to decide on steps and may call tools to carry them out. Our guide to agent skills expands on reusable procedures. Installing a skill or writing a long prompt does not automatically create another model.
What to check when choosing
A short summary, repeated classification task and complex problem may call for different tradeoffs. Define an acceptable result first, then check whether your chosen option delivers it at a reasonable cost and wait time. Provider positioning is a starting point, not a test result for your assignment.
Record the full model name, reasoning effort and application. “I used GPT‑5.6 Sol in Codex with medium effort” describes an experience much more clearly than “I used GPT.” Explore our AI tools to investigate the available options.
Explore this article’s models
Compare what matters to you
USD per million tokens · catalog rates
| Specification | GPT-6 Astra | GPT-5.6 Luna |
|---|---|---|
| Input | $10 | $0.2 |
| Output | $50 | $1.2 |
| Cache | $1 | $0.02 |
GPT-6 Astra · Source: openrouter.ai ↗ · Checked Sep 12, 2026
GPT-5.6 Luna · Source: openrouter.ai ↗ · Checked Sep 12, 2026
View full comparison →
