GPT-5.6 Sol vs Terra vs Luna — How the OpenAI Tiers Changed in September 2026

October 1 note: This is a historical comparison of the original GPT-6 Sol/Luna launch, not a current OpenAI routing recommendation. GPT-6.1 Sol has since launched; see its model profile and current API documentation for its ID, availability, and rates. The September launch figures below are retained as a dated snapshot and do not establish current relative quality.

Historical launch snapshot (checked 26 September 2026): On 22 September 2026 OpenAI launched GPT-6 Sol ($2 input / $10 output per million tokens) and GPT-6 Luna ($0.10 / $0.50). Those launch prices were 50% below the GPT-5.6 Sol and Luna promotional prices then in effect. The launch API IDs were gpt-6-sol and gpt-6-luna. GPT-6.1 Sol later launched on 29 September; consult its current profile before using any model ID or rate. This article records the September comparison and does not assert that GPT-5.6 Terra is the only current middle tier.

OpenAI's model family is best understood as a tier routing problem, not a "which model is smartest" problem. The profiles linked below record the respective models and dated launch information; newer versions and account availability can change which tier is appropriate. Match task shape to a model you can actually access, then validate it on your own prompts.

What each tier costs now

In that September comparison, GPT-6 Sol listed at the same input price as GPT-5.6 Terra and a lower output price. This was a unit-price comparison, not evidence of comparable quality or a present-day routing default. Confirm current model versions, account access, and rates before changing production traffic.

GPT-6 Sol is the frontier workhorse. OpenAI says it makes about half as many factual mistakes as GPT-5.6 Sol on its internal factuality evaluation. On FrontierCode it improves substantially over GPT-5.6 Sol. Those are vendor figures, so replay your own tasks before you trust them.

GPT-6 Luna is the volume tier. OpenAI says that at higher effort settings it matches GPT-5.6 Sol on the same factuality evaluation at about a hundredth of the cost. Pin the full model ID (gpt-6-luna) in configuration and billing dashboards before you interpret a cost spike. Aliases are how Luna and Sol traffic gets mixed up.

GPT-5.6 Terra keeps a narrow role: workloads already validated on Terra, where the migration test costs more than the saving.

GPT-6 Astra was described by OpenAI at launch as its strongest model for broad reasoning and computer use, at the launch price of $10/$50 per million tokens. Treat that as a vendor launch claim and verify current availability/pricing in the model profile before use.

A routing framework that survives repricing

Start with task class, not model curiosity.

This table is a starting point. Model prices move. Check the date on each row in the AI models index before committing routing logic.

When to step up or down

Step down from Sol to Luna when throughput matters more than frontier quality and your acceptance tests do not regress. Compare against non-OpenAI budget options too: Gemini 3.8 Flash lists at $0.75/$3.75 (intro pricing through 31 December 2026). The catalog also publishes Gemini 3.7 Flash vs GPT-5.6 Luna for the previous generation.

Step up from Sol to Astra when hard reasoning, long-horizon agent loops, or computer use fail on Sol in your pilot. Token rate is only half the economics. Retries and human review minutes are the other half.

Stay on a GPT-5.6 ID only where you have a validated pipeline and the re-test is not yet scheduled. Note that the GPT-5.6 Sol $4/$20 price is itself a promotion; the pricing cliffs post tracks its guaranteed period and any newly announced price change.

How to pilot without inventing benchmarks

Do not copy leaderboard scores into your routing rules. Use the catalog's verified fields (pricing, context, modalities) as inputs, then run your own replay:

1. Freeze 20–30 representative prompts across coding, synthesis, and structured output.
2. Route identical traffic to the old and new IDs behind the same adapter for two evaluation windows.
3. Score pass rate, review minutes, and total spend per successful task — not raw token volume.

The AI models index links every profile, comparison, and switch guide from one surface. If you are routing models inside a coding harness rather than a raw API, the AI coding task guide separates model choice from harness choice, which is a common source of tier mismatch.

The mistake that shows up in month two

Teams pick the flagship, then route everything through it, including traffic the volume tier would have passed. In September a second version of that mistake appeared: teams stay on GPT-5.6 IDs after a price cut and pay twice what they need to. The fix is the same both times. Put tier routing in application config, pin IDs, log which tier served each request, and promote a change only after pass rate holds for two windows.

If you want the trade-offs scored against autonomy and editor fit as well as model tier, the AI Coding Assistant Finder walks through both layers in one pass. For a model-only shortlist, use the AI Model Picker.

Sources

• OpenAI — Introducing GPT-6 Sol and Luna (22 Sep 2026: $2/$10 and $0.10/$0.50, 50% below GPT-5.6 promo pricing; gpt-6-sol, gpt-6-luna): https://openai.com/index/introducing-gpt-6-sol-and-luna/ — accessed 2026-09-26
• OpenAI — GPT-5.6 (Sol $4/$20 promo, Terra $2/$12, Luna $0.20/$1.20): https://openai.com/index/gpt-5-6/ — accessed 2026-08-29
• OpenAI API pricing: https://platform.openai.com/docs/pricing — accessed 2026-08-29
• OpenAI — Introducing GPT-6 Astra ($10/$50): https://openai.com/index/gpt-6-astra/ — accessed 2026-09-05
• TechCrunch — OpenAI launches GPT-6 Sol and Luna (22 Sep 2026): https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/ — accessed 2026-09-26