GPT-5.6 Explained: Sol, Terra, Luna and ChatGPT

GPT-5.6 is not one identical model sold at three arbitrary prices. OpenAI introduced durable capability tiers intended for different workloads. Choosing between them is an engineering decision about quality, latency, cost and risk.

This page records the current public information. Model access and product defaults can change, so verify the linked OpenAI documentation before making a procurement or production decision.

The GPT-5.6 family

Where GPT-5.6 is available

OpenAI’s July 2026 launch made the family available across ChatGPT, Codex and the API, but the exact tier and control exposed to a user depend on the product.

How to choose a tier

Start with the consequence of failure

A formatting task and a security-sensitive code review should not use the same evaluation threshold. Define what an unacceptable result looks like before selecting a model.

Test representative work

Build an evaluation set from real, authorized examples. Include ordinary cases, difficult cases and known failures. Compare task success rather than judging a few polished demonstrations.

Measure total system cost

Token price is only one factor. Retries, long outputs, tool calls, human review and failed automation can cost more than the original inference. A more capable model may be cheaper if it succeeds in fewer attempts. A smaller model may win when the task is constrained and verifiable.

Route work deliberately

A production system can send straightforward requests to a lower-cost tier and escalate difficult cases. Routing must be tested. An unreliable complexity detector can silently send high-risk work to the wrong path.

A practical selection matrix

This table is a starting hypothesis, not a benchmark result. Your own evaluation set should decide the production tier.

A practical evaluation design

Suppose a team wants a model to turn support conversations into structured incident summaries.

1. Define the required JSON schema and evidence fields.
2. Remove or protect personal information before testing.
3. Create examples covering simple tickets, ambiguous tickets and multi-issue conversations.
4. Score schema validity, factual support, omitted details and invented details separately.
5. Run the same examples against Sol, Terra and Luna with controlled settings.
6. Measure latency and complete workflow cost.
7. Choose the least expensive tier that meets the agreed quality and safety threshold.
8. Monitor live failures and keep a rollback path.

This process produces a defensible decision. Asking which model is “best” without a task and threshold does not.

GPT-5.6 is still a model, not the complete application

A secure product also needs authentication, authorization, data handling rules, prompt construction, retrieval controls, tool permissions, output validation, monitoring and human escalation. Model capability cannot replace those boundaries.

When the model uses a tool, the application should validate arguments and authorize the action. Do not grant access merely because a generated tool call looks confident.

Current limitations to plan for

What to record in production

• model ID and relevant configuration;
• prompt or instruction version;
• tool calls and validation outcomes;
• latency and token usage;
• automated evaluation signals;
• user correction or escalation;
• privacy-safe failure examples for regression testing.

If the distinction between model and product is still unclear, return to GPT vs ChatGPT. Developers should also consult the current OpenAI model directory before implementation.