GPT-5.6 Sol, Terra, Luna Compared: Pricing, Capabilities, and Use Cases

Fri, 10 Jul 2026 13:26:11 +0800

On July 9, 2026, OpenAI officially launched the GPT-5.6 family: flagship Sol, balanced Terra, and low-cost Luna. They share the same generation, but differ in capability, speed, price, and best-fit tasks.

The simplest rule is: use Sol for difficult work, Terra for most everyday work, and Luna for high-volume lightweight jobs. Sol is not automatically the best choice for every request; Terra and Luna are usually better value when results are easy to check.

Core differences between Sol, Terra, and Luna

Model	Positioning	API input price	API output price	Best for
GPT-5.6 Sol	Flagship, highest capability	$5 / million tokens	$30 / million tokens	Complex coding, long-running agents, professional research, hard problems
GPT-5.6 Terra	Balanced capability and cost	$2.50 / million tokens	$15 / million tokens	Everyday work, routine coding, content processing, default workflows
GPT-5.6 Luna	Fastest and cheapest	$1 / million tokens	$6 / million tokens	Classification, extraction, summaries, batch and high-throughput tasks

The number 5.6 identifies the model generation; Sol, Terra, and Luna are durable capability tiers. OpenAI says each tier can evolve on its own schedule instead of waiting for the entire family to be replaced at once.

GPT-5.6 Sol: for the hardest problems

Sol is the flagship and is suited to tasks where correctness and completeness matter more than call cost. It is strongest at complex coding, long-running professional work, cybersecurity, and scientific research, and is better at using tools, operating computers, and checking final results.

Typical examples:

Modify several modules in a large codebase, run tests, locate failures, and continue fixing them.
Combine many files and online sources into a deliverable report, presentation, or spreadsheet.
Agent workflows that run many tools while retaining long-range context.
Problems where repeated failures by ordinary models would make manual rework expensive.

Sol’s API output price is twice Terra’s and five times Luna’s. Using it for simple summaries, format conversion, or classification rarely justifies the cost.

GPT-5.6 Terra: the default for most people

Terra sits between the other two in capability, speed, and price. OpenAI positions it as the balanced model for everyday work. It is a better default entry point for ordinary conversations, office writing, routine programming, information organization, and most tool calls.

Terra works well for:

Writing routine scripts, fixing well-scoped bugs, and explaining code.
Organizing documents, writing email, drafting content, and summarizing meetings.
AI features that need stable quality at relatively high call volume.
A tiered workflow that starts with Terra and upgrades difficult cases to Sol.

API developers can establish a quality baseline with Terra. Move requests down to Luna if evaluations show it is reliable; route complex samples to Sol if error rates remain high.

GPT-5.6 Luna: cheap, fast, and built for volume

Luna is the fastest and least expensive tier. It fits tasks with clear steps, easy validation, low per-call value, and many repetitions.

Common uses include:

Text classification, label generation, and key-value extraction.
Short summaries, title candidates, and format cleanup.
Batch processing of support records, product data, or logs.
First-pass filtering and simple subtasks inside agent workflows.

Luna should not be the only model for high-risk decisions or complex engineering. A safer design automatically escalates low-confidence, schema-invalid, or repeatedly retried requests to Terra or Sol.

`max` and `ultra` are not new model tiers

max and ultra are often discussed alongside Sol, Terra, and Luna, but they are not a fourth or fifth model.

max means higher reasoning effort. It gives the model more time than xhigh to explore approaches, run checks, and revise an answer. It suits difficult tasks, at the cost of more tokens and potentially longer latency.

ultra is a multi-agent mode that coordinates four agents in parallel and combines their results. It fits complex research and engineering tasks that can be split into independent branches. It only adds overhead for simple questions.

The distinction is:

Choose Sol, Terra, or Luna to select the model tier.
Choose reasoning effort to decide how much compute one task receives.
Consider ultra only when parallel decomposition genuinely helps.

Which models can ChatGPT and Codex users access?

GPT-5.6 began rolling out globally on July 9, with the full rollout expected within 24 hours.

ChatGPT: Plus, Pro, Business, and Enterprise users can use Sol with medium or higher reasoning effort; Pro and Enterprise can also choose Sol Pro.
ChatGPT Work and Codex: Free and Go users can use Terra; Plus, Pro, Business, and Enterprise users can choose Sol, Terra, or Luna.
max: In ChatGPT Work and Codex, anyone with GPT-5.6 access can enable it in settings.
ultra: ChatGPT Work offers it to Pro and Enterprise; Codex offers it to Plus and higher plans.
OpenAI API: Developers can call Sol, Terra, and Luna.

If an option is missing, the staged rollout may not have reached the account yet.

How API developers can combine the three models

A practical design is a tiered router rather than one fixed model:

Use Luna for classification, extraction, cleanup, and high-throughput preprocessing.
Use Terra for default conversations, ordinary generation, and routine coding.
Escalate difficult, low-confidence, or repeatedly failed requests to Sol.
Increase reasoning effort for high-value complex work, using max when necessary.
Use ultra only when work can be parallelized and time matters more than token cost.

GPT-5.6 also supports explicit prompt-cache breakpoints and a cache lifetime of at least 30 minutes. Cache writes cost 1.25 times the uncached input price, while cache reads retain a 90% input-price discount. Reuse long system prompts, fixed tool definitions, and repeated background context whenever possible.

One-line selection guide

Highest capability and less rework on valuable tasks: Sol.
A balance of quality and cost, or unsure where to start: Terra.
Simple, high-volume tasks with easy automated checks: Luna.
A single very difficult problem: increase reasoning effort or use max.
Work that splits into parallel workflows: consider ultra.

This site previously covered GPT-5.6 Sol’s limited preview tiers and safety strategy. After the official launch, the three tiers became selectable product levels in ChatGPT, Codex, and the API.

Original: GPT-5.6: Frontier intelligence that scales with ambition

AI模型 on KnightLi Blog