GPT-5.6 Sol vs Terra vs Luna: Which One Should You Use?
OpenAI's GPT-5.6 family comes in three tiers: Sol, Terra, and Luna.
Sol is the flagship. Terra sits in the middle. Luna is the fast, lower-cost option. That sounds simple, but the benchmark gaps are much smaller than the price gaps—and the cheapest tier has one weakness that is easy to miss.
The practical answer is:
- Start with Terra for most coding, analysis, and professional work.
- Choose Sol when a small quality advantage could materially change the result.
- Choose Luna for fast, repeatable work that does not rely on extremely long context.
One detail is worth knowing immediately: the plain gpt-5.6 alias points to Sol. If you use the default name without choosing a tier, you are choosing the flagship price too.
The three tiers at a glance
All three models offer about a 1.05-million-token input window, up to 128,000 output tokens, image and PDF understanding, reasoning, tool calling, and structured outputs on NanoGPT. Their token prices follow a consistent ratio:
| Model | Intended role | Relative token price | Best starting point for |
|---|---|---|---|
| GPT-5.6 Sol | Flagship | 5x | Hard reasoning, complex agents, demanding coding, high-stakes review |
| GPT-5.6 Terra | Balanced | 2.5x | Everyday coding, analysis, documents, and professional work |
| GPT-5.6 Luna | Fast and economical | 1x | High-volume chat, extraction, classification, and lightweight workflows |
Sol costs twice as much per token as Terra. Terra costs two and a half times as much as Luna. That does not mean Sol costs exactly five times as much per completed task: models can use different numbers of tokens to reach an answer.
Independent testing strengthens Terra's case
Artificial Analysis tested the three models at maximum reasoning effort through OpenAI's API. Its composite Intelligence Index and performance measurements show a clear tradeoff between quality, speed, and how many tokens each model uses.
| Model | Intelligence Index | Output speed | Evaluation output tokens |
|---|---|---|---|
| Sol | 59 | 54 tokens/second | 70 million |
| Terra | 55 | 145 tokens/second | 96 million |
| Luna | 51 | 199 tokens/second | 130 million |
Sol scores highest, as expected. But Terra gives up only four points while producing output about 2.7 times as quickly in this test. For interactive work, that difference in waiting time can matter more than a small benchmark lead.
Luna is faster again and has the lowest per-token price. It also used substantially more evaluation output tokens. Cheaper per-token pricing does not always translate to equally cheap completed tasks.
These are independent benchmark runs at maximum reasoning effort, not a promise about every NanoGPT conversation. The standard NanoGPT entries default to medium reasoning, so real results may use fewer tokens and show different gaps.
OpenAI's benchmarks show where the gaps widen
OpenAI's launch results broadly support the same picture. Terra stays close to Sol across coding and terminal tasks. Luna remains competitive on many ordinary benchmarks, but falls much further behind when the model must find and use information buried in an extremely large context.
| OpenAI-reported benchmark | Sol | Terra | Luna |
|---|---|---|---|
| Coding Agent Index | 80.0 | 77.4 | 74.6 |
| Terminal-Bench 2.1 | 88.8 | 87.4 | 84.7 |
| FrontierMath, tiers 1–3 | 89.0 | 84.9 | 78.6 |
| MRCR long-context retrieval, 512K–1M | 73.8 | 72.5 | 41.3 |
The final row is the important one. All three models accept very large inputs, but accepting a document and using it reliably are not the same thing. Terra nearly matches Sol when the relevant information is buried across roughly half a million to one million tokens. Luna does not.
These are OpenAI's own published numbers. They are useful for comparing the shape of the family, but they should still be treated as first-party launch results rather than a neutral guarantee for every workload.
Why Terra is the best default
Terra is the sensible starting point for most people because it avoids the biggest compromise on either side.
It is strong enough for:
- Everyday coding and debugging
- Reviewing documents and spreadsheets
- Research and comparison work
- Tool-using workflows
- Long conversations and large project context
- Drafting polished professional material
In the published results, Terra is close to Sol on coding, terminal work, and very long context. It also streamed output much faster than Sol in independent testing while costing half as much per token.
That does not make Terra universally better. It means the burden of proof should be on Sol: start with Terra, then move up only when your own task shows a meaningful improvement.
When Sol is worth the extra cost
Use Sol when the last few points of quality have real value.
Good examples include:
- A difficult software problem that may take many steps to resolve
- High-stakes analysis where missing one issue would be expensive
- Hard mathematics or scientific reasoning
- Long autonomous agent runs where small mistakes can compound
- A final review of work produced by a cheaper model
Sol has the highest long-context score in the family, but Terra is only 1.3 points behind on the very-long-context result above. The meaningful divide is Sol and Terra versus Luna. Do not pay the Sol premium for context length alone; choose it when that large context is combined with unusually difficult reasoning, coding, or review work.
Remember that the plain gpt-5.6 model name routes to Sol. That is a reasonable quality-first default, but it is worth choosing Terra explicitly when the flagship margin is unnecessary.
When Luna is the right tool
Luna makes sense when you need many answers quickly and the work is fairly repeatable:
- Classification and tagging
- Extracting fields from text
- Rewriting, translation, and formatting
- Short customer-service responses
- Generating many variants
- Lightweight coding help
- Fast chat where latency matters
It can still reason, use tools, and accept large inputs. The benchmarks simply suggest that its practical reliability drops when the task depends on retrieving details from the far reaches of a very large context.
Luna is therefore a strong throughput model, not a drop-in replacement for Sol on every task. Keep prompts focused, watch output length, and test it before moving a large long-context workflow.
Reasoning effort also matters
NanoGPT's standard GPT-5.6 entries default to medium reasoning and let you choose none, minimal, low, medium, high, or xhigh. The separate Pro entries add max and default to it.
Before moving from Terra to Sol, try the same task at a higher Terra reasoning level. Before moving from Luna to Terra for a simple workload, check whether Luna at low or medium already produces the result you need.
More reasoning can improve difficult answers, but it also adds time and may use more output tokens. When comparing tiers, test the same effort first so the comparison is fair. OpenAI's migration guidance also recommends testing your current reasoning setting and one level lower because GPT-5.6 may preserve quality with less effort.
A simple selection rule
If you do not want to benchmark every prompt:
- Use Terra as the general default.
- Move to Sol for genuinely difficult or consequential work.
- Move to Luna for high-volume, latency-sensitive, shorter-context tasks.
- Test a few real prompts before changing a production workflow.
You can try GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna on NanoGPT.
