GPT-5.6 Sol vs Terra vs Luna: Differences, Pricing and Which to Use

Published on July 18, 2026 by Steven Jones

GPT-5.6 Sol, Terra and Luna are three capability tiers within the same OpenAI model generation. Sol is the most capable, Terra balances performance with cost, and Luna prioritises speed and lower-cost processing.

The difficult part is not remembering that order. It is deciding whether a task genuinely needs a stronger model, more reasoning effort, or simply a better prompt. This comparison explains the practical differences across ChatGPT, Codex and the OpenAI API, including pricing, availability, workload fit and the hidden cost of choosing more capability than the job requires.

Model	Position in the range	Best suited to	API input price	API output price
GPT-5.6 Sol	Flagship	Difficult reasoning, complex coding, research, computer use and long-running agent work	$5 per 1 million tokens	$30 per 1 million tokens
GPT-5.6 Terra	Balanced	Everyday development, professional analysis and workflows requiring strong results at lower cost	$2.50 per 1 million tokens	$15 per 1 million tokens
GPT-5.6 Luna	Fastest and cheapest	High-volume processing, extraction, classification, routine automation and simpler coding tasks	$1 per 1 million tokens	$6 per 1 million tokens

API prices above are based on OpenAI’s published rates for standard uncached text tokens. Tool charges, cache writes and other product costs may be additional.

Sol, Terra and Luna are capability tiers, not reasoning settings

The first distinction prevents most model-selection mistakes: Sol, Terra and Luna identify the model tier, while Medium, High, Extra High, Max and Ultra describe how much reasoning or orchestration the selected product allows for a task.

A user can therefore face a choice between a stronger model that uses modest reasoning and a cheaper model that uses greater reasoning effort. Those options are not interchangeable. Increasing effort gives a model more opportunity to inspect, revise or explore a problem, but it does not give Luna the same underlying capability profile as Sol.

This creates a useful decision rule. Change the reasoning level when the model understands the type of problem but needs more time to work through it. Move to a stronger tier when it repeatedly misunderstands the task, overlooks dependencies, makes poor strategic choices or cannot maintain quality across a long workflow.

OpenAI describes Sol, Terra and Luna as durable tiers that can advance on separate schedules. The version number identifies the generation, while the tier names indicate their relative position. A future Terra update could therefore arrive without every tier changing at exactly the same time.

---

---

GPT-5.6 Sol is for work where failure creates expensive rework

Sol is OpenAI’s flagship GPT-5.6 model. It is the appropriate starting point for tasks where the model must coordinate several difficult activities rather than produce a single plausible response.

Examples include tracing a fault across a large codebase, planning a migration with several dependencies, comparing conflicting research sources, controlling software through computer use, producing a polished client deliverable or running an agent that must recover when tools fail.

Sol’s advantage is most visible when mistakes compound. A weaker model may complete each individual step reasonably well but choose the wrong sequence, lose track of an earlier constraint or accept an incomplete result from a tool. Sol gives you a better chance of maintaining the objective across the complete workflow.

That does not mean Sol should be the default for every prompt. Using it to classify support tickets, clean simple text fields or summarise short documents wastes money without guaranteeing a materially better business result. It can also feel slower because the product may allocate more work to reasoning and verification.

GPT-5.6 Terra is the practical default, but not automatically the best value

Terra sits between Sol and Luna. OpenAI positions it as the balanced option for everyday work, with performance competitive with GPT-5.5 at a lower API price.

It is a sensible default for feature development, document analysis, structured research, data interpretation, routine agent workflows, and business writing that needs more judgement than Luna can reliably supply. Teams moving from a single-premium model can use Terra as the baseline and escalate only the tasks that fail the defined quality checks.

The hidden limitation is that a middle tier is not always the economic winner. Terra costs two and a half times as much as Luna per token. If Luna completes a predictable extraction task with the same acceptance rate, Terra adds cost without adding useful value. At the opposite end, a difficult task that needs several Terra retries may cost more than one successful Sol run.

Do not choose Terra simply because it sounds safer than Luna and cheaper than Sol. Give it a role in a measured workflow: define the output requirements, test failure rates and compare the cost per accepted result.

GPT-5.6 Luna is built for throughput, not just short answers

Luna is the fastest and least expensive GPT-5.6 model. OpenAI describes it as suitable for cost-sensitive, high-volume workloads, roughly corresponding to the nano tier used in earlier GPT-5 model families.

Good Luna workloads tend to have clear boundaries and outputs that are easy to validate. Classification, entity extraction, metadata generation, document routing, first-pass summaries, standardised transformations and repetitive code changes all fit that pattern.

Luna also supports reasoning, tool use and a large context window. That makes it more capable than the phrase “cheap model” suggests. The operational constraint is predictability. Giving Luna one million tokens of poorly organised context does not ensure it will identify the same subtle dependencies as Sol.

A recurring real-world mistake is to increase Luna’s reasoning setting until it becomes a substitute for every other tier. That can work on isolated examples, but it needs measurement. Higher effort can increase latency and consumption while still leaving a capability gap on difficult planning, scientific reasoning or autonomous software work.

The API price gap becomes significant once output grows

Output tokens cost six times more than input tokens across all three GPT-5.6 tiers. Long-generated reports, verbose agent traces, and repeated code rewrites can therefore dominate the bill.

Example monthly usage	Sol	Terra	Luna
10 million input tokens	$50	$25	$10
2 million output tokens	$60	$30	$12
Combined model cost	$110	$55	$22

These figures exclude tool calls, cache-write charges, storage and retrieval infrastructure, and failed requests. Our guide to AI hosting costs explains why the model’s token price is only one part of a production application’s budget.

The cheapest model per token can also be the most expensive per accepted result. Suppose Luna needs three attempts to produce a valid output while Terra succeeds once. Luna’s headline rate is lower, but retries consume tokens, add latency and complicate application logic. Measure the acceptance rate, not just the invoice rate printed next to the model name.

Which GPT-5.6 models can you select in ChatGPT?

Standard ChatGPT conversations do not present Sol, Terra and Luna as three equivalent buttons. Eligible paid users can access GPT-5.6 Sol by making reasoning choices such as Medium or High. Extra High and Sol Pro depend on the subscription plan.

Terra and Luna are not selectable in ordinary ChatGPT conversations. OpenAI makes the three tiers available in other environments, including ChatGPT Work, Codex and the OpenAI API, with access varying by plan.

This product distinction explains why two users can discuss GPT-5.6 but see different model controls. One may be using a normal chat, another Codex, and another an API application with a model identifier such as gpt-5.6-terra.

A task-routing framework is better than choosing one permanent model

The most efficient implementation does not send every request to the same tier. It routes work according to uncertainty, consequence and validation difficulty.

Task characteristic	Start with	Escalate when
Repetitive, structured and easy to validate	Luna	The invalid-output or exception rate exceeds your threshold
Requires judgement but follows a familiar workflow	Terra	Retries reveal missed context, weak planning or unstable tool use
Open-ended, high-consequence or difficult to verify	Sol	Increase reasoning effort before redesigning the workflow
Large batch with a small number of difficult exceptions	Luna with escalation	Confidence checks flag uncertain or malformed results
Long-running agent with several tools and dependencies	Sol or Terra after evaluation	The model loses objectives, mishandles permissions or fails recovery tests

A strong production pattern uses Luna for the first pass, deterministic checks to identify failures, and Terra or Sol only for exceptions. This preserves premium capability for the requests that benefit from it.

Routing should not rely solely on a model’s self-reported confidence. Validate outputs using schemas, tests, business rules, source checks or human review. A fluent answer is not evidence that the selected tier handled the task correctly.

How to evaluate Sol, Terra and Luna on your own workload

Public benchmarks help establish the broad capability order, but they cannot tell you which model delivers the lowest cost per usable result in your application. Build a small evaluation set from real work instead.

1. Select 30 to 100 representative tasks, including straightforward requests, edge cases and known failures.
2. Run the same prompts, context and tool permissions against all three tiers.
3. Score correctness, instruction adherence, tool success, latency, token use and the amount of human correction required.
4. Repeat a sample several times to expose inconsistent behaviour rather than judging one attractive result.
5. Calculate cost per accepted output and cost per completed workflow, not only cost per request.
6. Set escalation rules for tasks that fail validation or exceed an uncertainty threshold.

Keep the reasoning setting fixed during the first comparison. Otherwise, you cannot tell whether a difference came from the model tier or the amount of reasoning allocated. Test alternative effort settings in a second round.

For wider comparisons across text and language model products, see the DIY AI AI text-generation tools dataset.

Common GPT-5.6 model-selection mistakes

Using Sol before improving the prompt

A premium model cannot reliably repair missing requirements, contradictory instructions or disorganised context. Improve the task definition before paying for additional capability.

Comparing models with different tools or permissions

A model with repository access, web search or a working MCP connection will appear more capable than one tested without those resources. Keep the environment consistent.

Treating maximum reasoning as a free quality upgrade

More reasoning can improve difficult results, but it may also raise latency and usage. Apply it to tasks where additional investigation has a realistic chance of changing the answer.

Ignoring the cost of review and retries

A lower API bill is not a saving if staff spend longer correcting output. Include review time, failure handling and reruns in the comparison.

GPT-5.6 Sol vs Terra vs Luna verdict

Choose Luna for predictable, high-volume work with strong validation. Choose Terra for everyday development and professional tasks that need judgement without Sol’s full price. Choose Sol for difficult, high-consequence workflows where weak planning or missed dependencies would create more cost than the model upgrade.

For individual use of ChatGPT, the decision may appear to be a reasoning-level choice because standard chats primarily expose GPT-5.6 Sol rather than all three tiers. Developers using Codex, Work, or the API have the more interesting option: route ordinary requests to Luna or Terra and reserve Sol for the exceptions that justify it.

The best default is not one model. It is a controlled escalation path backed by your own evaluation set.

Frequently asked questions

Is GPT-5.6 Sol better than Terra?

Sol is the more capable flagship tier, particularly for complex reasoning, long-running workflows, computer use and difficult coding. Terra can still be the better operational choice when it meets the required quality level at half Sol’s token price.

Is GPT-5.6 Luna good for coding?

Luna can handle routine coding, focused edits, code explanation and well-scoped automation. Use Terra or Sol when a task requires architectural decisions, extensive repository knowledge, difficult debugging, or coordination across several tools.

Can I select Terra or Luna in a normal ChatGPT conversation?

No. OpenAI says Terra and Luna are not selectable in standard ChatGPT conversations. They are available in products such as ChatGPT Work, Codex and the OpenAI API, subject to plan access.

Do all GPT-5.6 models support reasoning?

Yes. The tiers support reasoning, but the underlying model and the selected reasoning effort are separate controls. Raising Luna’s effort does not automatically make it equivalent to Sol.

Where can I verify the current GPT-5.6 pricing?

OpenAI publishes current availability, pricing and product access in its official GPT-5.6 announcement. Check the live documentation before committing pricing to a production forecast, as model access and commercial terms can change.