GPT-5.6 Sol vs Terra vs Luna: Which to Use

The most expensive mistake we see in Malaysian AI deployments has nothing to do with prompting. It is running every request through the flagship model because someone picked it once during a proof of concept and nobody revisited the decision.

OpenAI made that mistake easy to fix and easy to make. With GPT-5.6 the company stopped shipping a single model per generation. The number identifies the generation; Sol, Terra and Luna identify durable capability tiers that can advance independently. GPT-6 Astra now sits above all three, but Sol, Terra and Luna remain available in the API and in ChatGPT — and for most production workloads they are still the right answer.

1. The three tiers, plainly

All three share the same envelope: a 1,050,000-token context window, 922,000 maximum input tokens, 128,000 maximum output tokens, and a 16 February 2026 knowledge cutoff. All three support the same reasoning-effort ladder — none, low, medium (default), high, xhigh, max — and the same tool set: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search.

What differs is capability and price.

• Sol — the flagship. Deepest reasoning, the only tier that unlocks ultra mode (which spawns subagents on hard problems). Pronounced gains in cybersecurity and biology. US$4 input / US$20 output per million tokens; cached input US$0.40.
• Terra — the balanced middle. Roughly competitive with the previous generation's flagship at about half the cost. US$2 input / US$12 output per million; cached input US$0.20.
• Luna — fast and cheap, built for volume and latency-sensitive work. US$0.20 input / US$1.20 output per million; cached input US$0.02.

Read those numbers again. Sol costs 20 times what Luna costs on input and roughly 17 times on output. That spread is the entire argument for taking model selection seriously.

2. Where the tiers actually separate

Benchmarks split them less cleanly than the pricing suggests. On Terminal-Bench 2.1, Sol in ultra mode scored 91.9%, Sol alone 88.8%, Luna 84.3% and Terra 82.5% — note that Luna scored above Terra on that particular benchmark, which should make you suspicious of any simple "more expensive is better" heuristic.

The separation shows up on long-horizon work. CodeRabbit's benchmark across 100+ multi-language coding tasks put Sol at a 63.7% pass rate against Terra's 40.7%. More telling: Terra burned an average of 55,594 output tokens per task versus Sol's 20,968. Terra thought harder and got there less often.

On code review the same pattern held — Sol at 69.7% pass rate, Terra at 52.5%. Terra's actionable precision was actually higher (35.7% vs 31.6%), but it produced 143 raw comments to Sol's 231. Terra is quieter and more selective. Sol finds more and needs filtering.

The lesson is the one we keep repeating in our AI automation ROI work: the meaningful unit is cost per solved task. Terra at half price that fails twice as often is not a saving.

3. A routing map that works

Here is the allocation we recommend to teams starting from scratch. Route by task type, not by department.

4. The reasoning-effort dial nobody uses

Model choice gets all the attention; reasoning effort gets almost none. Every tier exposes six settings from none through max, and the cost difference between them on a long task can exceed the cost difference between tiers.

Practical defaults we use:

• none or low for extraction, classification, formatting, and anything with a deterministic right answer.
• medium — the default — for drafting, summarising, and first-pass analysis.
• high or xhigh for multi-step reasoning where a wrong answer is expensive to catch.
• max, and Sol's ultra mode, reserved for genuinely hard problems that a human would need half a day on.

Turning effort down on Sol is frequently cheaper and better than turning effort up on Luna. Test it on your own workload before assuming otherwise.

5. Caching is the biggest lever, and it is free

Cached input costs 10% of standard input across all three tiers — US$0.40 on Sol, US$0.20 on Terra, US$0.02 on Luna. Prompt caching carries a 30-minute minimum life, with cache writes billed at 1.25x the input rate.

For any application with a stable system prompt, a fixed document set, or a codebase the agent re-reads — which is to say most production applications — structuring prompts so the invariant part comes first turns a 10x cost reduction into an architectural decision rather than a negotiation. Teams that get this right routinely cut their bill by 60–80% without changing model. Our explainer on tokens and tokenisation covers the mechanics.

6. Should you move to Astra instead?

GPT-6 Astra shipped on 3 September 2026 at US$10 input / US$50 output — 2.5x Sol. It is materially better at computer use, long-horizon agents and security work. On general knowledge work, independent benchmarking found it roughly level with Sol, at 75% higher cost per task.

So the honest answer for most Malaysian organisations is: Terra for the bulk of production work, Luna for high-volume classification and routing, Sol for the hard tail, and Astra only where the workload is genuinely agentic. We go into the trade-off in detail in our GPT-6 Astra breakdown.

Pick the cheapest tier that passes your evaluation set. Then measure again in a quarter, because OpenAI has now shown it will re-price a tier without renaming it.