GPT-5.6 Explained: How to Choose Sol, Terra, Luna, and the Right Reasoning Effort

“Use the fastest and least expensive configuration that can do the work reliably.”

Every time a new model arrives, the temptation is the same: pick the most powerful option, turn every setting to maximum, and assume the output must be better. In practice, that habit mostly buys you slower answers at a higher price.

I look at AI from a practical business and operating perspective. The important question isn’t “Which model is the smartest?” It’s “Which model fits this task, and what does the extra time and cost buy me?”

GPT-5.6 is a family, not a single model with one quality setting: three tiers (Sol, Terra, and Luna) plus reasoning controls that decide how much computational effort the model can spend on a task.

These tiers aren’t ChatGPT subscription plans. Sol, Terra, and Luna are models; Free, Go, Plus, Pro, Business, and Enterprise are plans, and your plan decides which models you can access.

OpenAI released GPT-5.6 on July 9, 2026. Access varies between standard ChatGPT conversations, ChatGPT Work, Codex, and the OpenAI API, and the initial rollout is gradual.

Official references: OpenAI announcement | GPT-5.6 availability in ChatGPT

The three GPT-5.6 tiers at a glance

The positioning and prices below come from OpenAI’s documentation. The task recommendations are mine, based on where each model has made sense in my own work.

Model	Official position	Where it fits	Avoid when	Cost and recommendation
GPT-5.6 Sol	The flagship GPT-5.6 model for complex professional work. The gpt-5.6 API alias points to Sol.	Difficult analysis, deep research, complex coding, major documents, design, ambiguous problems, and demanding reviews.	Routine rewriting, classification, extraction, or large volumes of simple work.	$5 input / $30 output per million standard short-context API tokens. Use Sol when quality matters more than the quickest response.
GPT-5.6 Terra	A balanced model combining intelligence and cost efficiency. It broadly occupies the earlier mini position.	Business writing, everyday analysis, ordinary coding, document review, research synthesis, and most professional tasks.	The hardest open-ended problems, or simple repetitive work that Luna can handle.	$2.50 input / $15 output per million standard short-context API tokens. My default for most professional work when available.
GPT-5.6 Luna	The fastest and lowest-cost GPT-5.6 tier, designed for cost-sensitive, high-volume workloads.	Extraction, classification, tagging, transformation, routing, and clear repeatable work.	Nuanced analysis, difficult research, architecture decisions, or consequential final review.	$1 input / $6 output per million standard short-context API tokens. Use it when a correct result is clearly defined.

Official references: Sol model documentation | Terra model documentation | Luna model documentation | API pricing

The prices above are API prices, not ChatGPT subscription fees. Long-context requests, regional processing, caching, and certain tools can have different charges.

One complication worth knowing: standard ChatGPT doesn’t present the complete family the way Codex or the API do.

In ordinary ChatGPT conversations, GPT-5.5 Instant remains the default for quick everyday responses. On eligible plans, GPT-5.6 Sol powers the Medium, High, and Extra High reasoning choices. Terra and Luna are not currently selectable in standard ChatGPT conversations, although they are available through supported ChatGPT Work and Codex experiences and through the API.

You may also see GPT-5.6 Sol Pro. That isn’t a fourth tier; it’s a higher-compute way of running Sol for particularly difficult work.

Official references: GPT-5.6 in ChatGPT

Which tier should I choose?

Start by asking how difficult the task actually is.

If the work is quick, disposable, and easy to check, use the fastest option available. In standard ChatGPT, that may mean staying with Instant rather than deliberately selecting GPT-5.6. You don’t need a flagship reasoning model to correct a sentence, format a list, or answer a basic question.

If the task requires analysis but the objective is clear, Terra at medium effort is a sensible starting point when you have access to it.

Move to Sol when the problem is ambiguous, high-value, technically difficult, or dependent on judgment across several sources.

• Quick everyday work: Use Instant in standard ChatGPT, or Luna with low effort in an automated workflow. Good examples include extracting dates, cleaning meeting notes, categorizing feedback, or drafting a short reply.
• Business writing and analysis: Start with Terra at medium. Move to Sol for a board paper, policy comparison, investment argument, procurement evaluation, or strategy involving competing priorities.
• Research: Terra is fine for summarizing a defined collection of sources. Sol becomes worthwhile when sources conflict, evidence is incomplete, or the model must investigate and defend a conclusion.
• Coding: Terra should handle many ordinary implementations, tests, and focused changes. Sol is better suited to unfamiliar repositories, architectural decisions, difficult debugging, migrations, and security-sensitive work.
• Large-document work: Use Terra for structured extraction and summarization. Use Sol when the answer depends on connecting evidence scattered across multiple files. A large context window helps, but it doesn’t remove the need for a clear question or careful verification.
• High-stakes review: Use Sol with high or xhigh reasoning, then involve the appropriate human specialist. More model effort is useful here, but it doesn’t transfer accountability to the model.
• High-volume automation: Start with Luna. Route uncertain or failed cases to Terra or Sol instead of paying for maximum capability on every request.

This last approach matters for developers and organizations. Good AI architecture doesn’t use one expensive model for everything. It sends each task to the lowest tier that can meet the required quality, with a clear escalation path for exceptions.

Understanding reasoning effort

Think of reasoning effort as a budget for how much work the model is allowed to put into the problem.

At higher settings, the model has more opportunity to explore alternatives, plan several steps, check its work, use tools, and revise its approach. This usually means waiting longer and, in the API, potentially using more billable tokens.

None of that makes your prompt clearer, fills in missing evidence, or guarantees a correct answer.

If you give the model a badly framed question, maximum effort may simply produce a more elaborate response to the wrong question.

GPT-5.6 supports none, low, medium, high, xhigh, and max through supported API configurations. Product interfaces expose different subsets and sometimes use different labels. For example, none is available only through the API, not in the apps or on any subscription path; Codex and ChatGPT Work use Light rather than low; and xhigh may appear as Extra High.

Official references: GPT-5.6 model guidance | ChatGPT Work and Codex model guidance

Effort	Meaning	Good uses	Avoid when	Example and trade-off
none	Minimizes deliberate reasoning.	Classification, extraction, routing, and simple transformations.	The task requires judgment, planning, or connected steps.	Classify support tickets and return JSON. Fastest and normally least token-intensive; an API-only setting.
low	Limited reasoning while favoring speed.	Drafting, routine coding, straightforward analysis, and simple tool use.	The problem is ambiguous or difficult to verify.	Turn meeting notes into actions, owners, and deadlines. Modest increase in time and token use.
medium	Balances reasoning depth, reliability, speed, and cost.	Most professional writing, analysis, research, coding, and planning.	The task is extremely simple or repeatedly fails at this level.	Compare proposals against evaluation criteria. The documented GPT-5.6 API default.
high	More room for planning and checking.	Difficult coding, serious research, complex decisions, and multi-stage analysis.	You only need a quick draft or simple transformation.	Review a migration plan for failure modes and missing tests. Slower and potentially more expensive.
xhigh	Extended reasoning for long, demanding tasks.	Deep research, security review, difficult code review, and complex synthesis.	Speed matters or high already produces an acceptable result.	Audit a repository with evidence. Substantial latency and token use; often displayed as Extra High.
max	More time than xhigh to explore, check, and revise.	The hardest quality-first problems and unusually consequential technical work.	Everyday writing, routine research, or work already handled by xhigh.	Stress-test a release architecture. Highest single-model reasoning investment; not a normal standard ChatGPT choice.

OpenAI recommends medium as the balanced starting point, low for latency-sensitive work, and max for the hardest quality-first tasks. It also recommends testing settings on representative work rather than assuming that the highest setting must be best.

Official references: Reasoning documentation

My recommended default settings

• Everyday ChatGPT use: Leave automatic switching enabled and allow ChatGPT to move from Instant to Medium when more reasoning is needed.
• Professional knowledge work: Start at medium for reports, presentations, document analysis, planning, and serious business writing.
• Writing and editing: Use low for clean-up and rewrites. Use medium when the structure or argument needs work. High is rarely necessary unless the model must reconcile evidence or challenge the content.
• Research: Begin at medium when the question and sources are clearly defined. Move to high when sources disagree or the conclusion requires several layers of analysis.
• Coding: Use Terra or Sol at medium for ordinary implementation. Use Sol at high for architecture, difficult debugging, major refactoring, or unfamiliar systems.
• API automation: Start with Luna at none or low. Measure accuracy, cost, and failure patterns. Escalate difficult cases instead of increasing the setting for the entire workload.

Developers should also pay attention to model IDs, token prices, latency, context size, prompt caching, tool costs, structured outputs, and evaluations. Ordinary ChatGPT users don’t need to manage most of this. Their main decisions are simpler: whether the task deserves GPT-5.6, and how much reasoning it deserves.

How to get better results without simply turning effort to max

A better prompt often creates more value than a higher reasoning setting.

1. Give the model the relevant context. Explain the situation, audience, available information, and why the output matters.
2. Describe the actual deliverable. Ask for a board memo, comparison table, recommendation, code patch, checklist, or executive summary. “Analyze this” is rarely enough.
3. State the constraints: length, tone, exclusions, deadline, permitted tools, any approval boundaries.
4. Set clear source rules. Say whether the model should browse, use only supplied documents, prefer primary sources, cite factual claims, or separate evidence from inference.
5. Provide a useful example. A short sample or reference document can communicate the expected quality better than several paragraphs of abstract instruction.
6. Explain how the work will be reviewed. Accuracy, completeness, readability, compliance, test coverage, and decision usefulness are different standards.
7. Iterate with a diagnosis. Don’t just say, “Try again.” Identify what was weak, what was missing, and what the next version needs to correct.

When a result is disappointing, I improve the assignment before I touch the effort setting. Higher effort lets the model do more work; it can’t tell the model what the right work is.

When not to trust the model alone

OpenAI advises users to evaluate outputs for accuracy and appropriateness and to apply human review, particularly in high-stakes settings.

Official references: OpenAI Terms of Use | API safety best practices

For consequential work, check the original sources, recalculate important figures, inspect citations, and ask the model to identify uncertainty or missing evidence.

Legal, medical, financial, safety-critical, and major business decisions should still involve qualified professionals and accountable decision-makers.

None of this is an argument against using AI for serious work, only for using it properly. The model can research, compare, draft, challenge, and review; what it can’t do is carry responsibility. That stays with the person or organization acting on the output.

A practical cheat sheet

If your task is…	Start with…	Move higher when…
A quick question or simple rewrite	ChatGPT Instant	The task develops into real analysis
Extraction, tagging, or classification	Luna, none or low	Exceptions require judgment
Routine business writing	Terra, low or medium	The argument is complex or consequential
A professional report or business analysis	Terra, medium	Evidence conflicts or the decision has high value
Research from supplied sources	Terra, medium	Sources disagree or important evidence is missing
Deep external research	Sol, high	A targeted revision still leaves major gaps
Routine coding	Terra, medium	The repository or change is complex
Difficult debugging or architecture	Sol, high	Several credible attempts have failed
Security or high-stakes technical review	Sol, xhigh	Max demonstrates a genuine improvement on comparable work
High-volume API processing	Luna, none or low	Confidence checks identify an exceptional case
A consequential decision	Sol, high, plus human review	Verification is always required before acting

Start at medium when the task genuinely requires reasoning. Increase the effort when you can explain what additional analysis or checking you expect to gain, not just because a higher setting exists.

Sources

GPT-5.6 official announcement https://openai.com/index/gpt-5-6/

GPT-5.6 in ChatGPT https://help.openai.com/en/articles/20001354-gpt-56-in-chatgpt

GPT-5.6 model guidance https://developers.openai.com/api/docs/guides/latest-model

Reasoning models and effort settings https://developers.openai.com/api/docs/guides/reasoning

GPT-5.6 API model documentation https://developers.openai.com/api/docs/models

OpenAI API pricing https://developers.openai.com/api/docs/pricing

ChatGPT Work and Codex model guidance https://learn.chatgpt.com/docs/models