Verified on August 19, 2026. Model availability, limits and pricing can change over time.

Update, July 30, 2026: OpenAI cut the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%, which also applies to credit metering in Codex and ChatGPT Work. The cost tables below already reflect the new values.

The official GPT-5.6 introduction video. Source: openai.com.

GPT-5.6 changed how you pick a model. Instead of one main option, OpenAI split the new family into three tiers: Sol, Terra and Luna.

Sol is the most capable model in the family. Terra balances quality, speed and cost. Luna was designed for faster, repetitive and price-sensitive work. OpenAI also introduced the Max reasoning level and the Ultra mode, which runs multiple agents in parallel.

That does not make Sol the right answer for everything. Running Sol with Max reasoning to write three product descriptions is like hiring a full crew to change a light bulb.

The right choice depends on four things:

  • How complex the task is.
  • How much quality it actually requires.
  • How much of it there is.
  • What a wrong answer costs.

The 30-second summary

Model or mode Best for Examples
GPT-5.6 Luna Clear, fast, repetitive, high-volume tasks Descriptions, titles, captions, classification, extraction and standardized summaries
GPT-5.6 Terra Everyday work that needs quality and context Articles, scripts, plans, proposals, emails and analysis
GPT-5.6 Sol Hard, ambiguous or high-value problems Deep research, strategy, decisions, critical review and polished deliverables
Max One hard problem that needs deeper reasoning Audits, investigation, diagnosis and complex planning
Ultra Large projects that split into separate tracks Multi-source research, full projects, parallel analysis and large deliverables
Pro Higher reliability and more internal work Very hard tasks where quality matters more than speed and spend

The official guidance positions Sol as the main model for complex professional work, Terra as the balanced option and Luna as the economical choice for large volumes.

Before you choose: where does each model show up?

One of the most confusing parts of the new family is that the three models do not appear together in every interface.

In regular ChatGPT, GPT-5.5 Instant is still the default model for fast answers. GPT-5.6 Sol powers the Medium, High and Extra High levels, while Sol Pro is used in the Pro option.

Terra and Luna cannot be selected in ordinary ChatGPT conversations. They are available in Work, in Codex and in the API, depending on the plan.

Product Current availability
Regular ChatGPT Sol on Medium, High and Extra High. Sol Pro on the Pro option
Work in ChatGPT Sol, Terra and Luna for Plus, Pro, Business and Enterprise
Codex Terra for Free and Go. Sol, Terra and Luna for Plus, Pro, Business and Enterprise
OpenAI API Sol, Terra and Luna
Max Available to users with GPT-5.6 access in Work and Codex
Ultra in Work Pro and Enterprise
Ultra in Codex Plus or above

GPT-5.6 Sol: for hard problems and high-value deliverables

GPT-5.6 Sol is the flagship of the family. It was built for complex work across programming, research, professional knowledge, science, security, computer use and design.

In practice, use Sol when the task is ambiguous or expensive to get wrong:

  • The request is incomplete or open to interpretation.
  • There are many constraints to reconcile.
  • The answer requires research, comparison or judgment.
  • A mistake causes loss, rework or a bad decision.
  • The output goes to clients, investors or an audience that matters.
  • Final quality outweighs a few extra seconds or credits.

Good uses for Sol outside programming

  • Researching and structuring an article from several sources.
  • Building an editorial strategy for a blog.
  • Analyzing different brand positions.
  • Reviewing an important commercial proposal.
  • Shaping the narrative of a presentation.
  • Comparing financial or operational scenarios.
  • Analyzing contracts (with professional review).
  • Developing the creative direction of a campaign.
  • Evaluating a complex user experience flow.
  • Planning an interactive/WebGL project.
  • Finding inconsistencies in a long document.
  • Reviewing critically for risks and gaps.

When Sol is usually overkill

  • Generating ten headlines from a formula.
  • Converting a list to JSON.
  • Summarizing a short, well-structured text.
  • Classifying comments by topic.
  • Rewriting hundreds of repetitive descriptions.
  • Fixing small grammar mistakes.

Sol is not simply "a smarter Luna". It earns its cost when there is exploration and judgment involved.

GPT-5.6 Terra: the model for everyday work

Terra was built to balance capability, speed and cost. It is the natural choice for tasks that need context, good writing and reasoning without demanding Sol's full depth.

For most teams, Terra is the right starting point.

Good uses for Terra

  • Writing a full article from a brief.
  • Creating a video script.
  • Preparing an email sequence.
  • Building a content calendar.
  • Structuring a commercial proposal.
  • Summarizing a meeting and organizing tasks.
  • Writing internal documentation.
  • Planning a campaign.
  • Analyzing a spreadsheet of ordinary complexity.
  • Creating a brief for design, video or 3D.
  • Editing a text while keeping the author's voice.
  • Turning scattered notes into an organized document.
  • Preparing a presentation from existing material.

When to move from Terra to Sol

  • Terra's output came out generic.
  • The sources conflict with each other.
  • The task requires spotting non-obvious risks.
  • The result carries real consequences.
  • The text needs rigorous editorial review.
  • There are many interdependent decisions.
  • Several attempts went nowhere.

When to move from Terra to Luna

  • The format is already defined.
  • The task repeats at high volume.
  • The text is short and objective.
  • There is a template to follow.
  • The result will be reviewed in a later step.

GPT-5.6 Luna: speed, volume and lower cost

Luna is the model optimized for cost-sensitive, high-volume work.

Good uses for Luna

  • Creating titles and subheadings.
  • Writing product descriptions.
  • Generating captions.
  • Producing variations of a hook.
  • Summarizing text in a fixed format.
  • Classifying messages and comments.
  • Extracting fields from documents.
  • Organizing unstructured data.
  • Turning text into lists, tables or JSON.
  • Standardizing file names.
  • Tagging assets.
  • Generating page metadata.
  • Adapting copy to different lengths.
  • Drafting a first version that will be reviewed.

In automation, a small difference in cost per call scales into real money across thousands of items.

Does Luna write well? Yes.

The better question is what kind of text it is and how much responsibility it carries.

Type of writing Recommended choice
Ten headlines for a post Luna
Short description from a fixed template Luna
Standardized summary Luna
Full article with context and voice Terra
Video script with a narrative Terra
An important email to a client Terra
Article with sources and deep synthesis Sol
Brand manifesto Sol
Critical review of an important document Sol
Ad variations at scale Luna to generate, Terra to review
High-value strategic content Terra to draft, Sol to review

Sol, Terra and Luna side by side

Criterion Sol Terra Luna
Reasoning capability Very high High Good
Speed Lower Balanced Higher
Cost Higher Medium Lower
Ambiguous tasks Excellent Good Limited
High volume Expensive Workable Ideal
Everyday writing Excellent, can overdo it Ideal Good for clear formats
Deep research Ideal Good for ordinary research Better for extraction and organization
Final polish Ideal Very good Basic
Repetitive automation Possible (more expensive) Good Ideal
Default choice for most work No Yes Depends on volume

Model, reasoning and orchestration are three different things

  • Model defines who does the task (Sol, Terra or Luna).
  • Max / reasoning level defines how much internal effort is applied.
  • Ultra defines when subagents work in parallel.

What is Max mode?

Max is the highest reasoning effort level. It returns answers with more exploration and revision, and it is usually slower.

Use Max to:

  • Investigate a hard problem.
  • Review an architecture.
  • Diagnose a persistent bug.
  • Compare decisions with many variables.
  • Analyze risk.
  • Plan a complex migration.

Do not use Max for simple, repetitive tasks.

What is Ultra mode?

Ultra splits a large task into parts and merges the results.

Use Ultra for:

  • Investigations with several tracks.
  • Competitive analysis.
  • Projects where subtasks are independent.

If the problem is tightly coupled, Max is usually the better fit.

What they consume and how to plan cost

To decide well, keep three things separate:

  • message limits in ChatGPT;
  • credits in Work/Codex;
  • and tokens in the API.

Credit table (illustrative estimate)

Model Input (credits/1M) Cached input Output
GPT-5.6 Sol 125 12.5 750
GPT-5.6 Terra 50 5 300
GPT-5.6 Luna 5 0.5 30

A practical example with 20k input tokens and 4k output tokens (no cache, no tools, short context):

Model Cost per task Cost of 100 tasks
Sol $0.22 $22.00
Terra $0.088 $8.80
Luna $0.0088 $0.88

At scale, the per-task difference becomes a financial decision.

A short answer can look cheap and still consume a lot, because of the internal reasoning work behind it.

API pricing (1M tokens)

Model Input Cached input Output
GPT-5.6 Sol $5.00 $0.50 $30.00
GPT-5.6 Terra $2.00 $0.20 $12.00
GPT-5.6 Luna $0.20 $0.02 $1.20

Values already include the July 30, 2026 reduction. The same announcement introduced Fast mode in the API (replacing Priority Processing): on Sol, up to 2.5x the speed at twice the price, with no change in intelligence.

How to plan usage without wasting resources

  1. Classify the task
  • Is it clear or ambiguous?
  • Will it run once or thousands of times?
  • Is a mistake easy to fix?
  • Does it require research?
  • Can it be split into independent parts?
  1. Start with the smallest model that can do it
  • Luna for volume and repetition.
  • Terra for everyday work.
  • Sol for hard or important tasks.
  1. Move up only when you need to
  • Luna organizes and transforms data.
  • Terra produces the deliverable.
  • Sol reviews the critical parts.
  • Max comes in for hard analysis.
  • Ultra comes in for parallelizable projects.
  1. Control the output
  • Define format, length, tone, criteria and what must not repeat.
  1. Do not send context you do not need
  • Oversized files, long history and irrelevant information all raise the bill.
  1. Use caching in repetitive flows

Where each model fits by area

Area Luna Terra Sol
Blog Titles, meta descriptions and tags Full article and calendar Research and final review
Marketing Ad variations Campaign planning Strategic positioning
Video Titles and captions Script and shot list Creative direction
Design File names and tags Briefs and documentation Visual direction and analysis
Research Extraction and organization Summary and comparison Synthesis and conclusion
Documents Standardization and cleanup Drafting and review Audit and refinement
Spreadsheets Classification and transformation Everyday analysis Scenarios and decisions
Customer support Standardized answers Contextual answers Critical cases

The support row is the one that fools people most. It is the pattern behind the ENE2ESE case, where the chatbot answers from the store catalog. Volume pushes you toward the cheap model, and the temptation is to buy back accuracy with more reasoning. It usually works the other way around: accuracy comes from querying the real data. That is how Nola AI, GUSTA's AI sales assistant answers price and stock questions in a store — a catalog query, not inference — which leaves the model handling only the conversation and keeps the cost per conversation low. Before picking a model, pick the layer: AI chatbot or AI agent for ecommerce.

An example workflow for articles

  • Luna: generate keywords, titles, FAQ and variations.
  • Terra: write the full text in the right tone of voice.
  • Sol: validate logic, risk and consistency.
  • Max: brought in when sources conflict or the stakes are high.
  • Ultra: useful on large projects with several separable tracks.

Quick answers

  • Which model for day-to-day work? Terra.
  • When should I pick Luna? Automation and volume in a predictable format.
  • When should I pick Sol? When there is ambiguity, risk and a business decision at stake.
  • Does Max always improve the result? No, sometimes it only raises the cost.
  • Is Ultra always better? No, it works best on parallelizable tasks.

Official sources and caveats

Check the rules below on the day you publish:

  • Availability can vary by plan and rollout phase.
  • Rates and limits can change.
  • Always confirm against OpenAI's current documentation.

Conclusion

GPT-5.6 does not require using the strongest model for everything.

A practical rule:

  • Luna for volume.
  • Terra for routine.
  • Sol for hard decisions.
  • Max for depth.
  • Ultra for parallelizable projects.

One line to decide by:

Luna for volume, Terra for the everyday, Sol for hard decisions, Max for depth and Ultra for projects with multiple tracks.