gpt-5.6 Routes to Sol, Not Luna — Avoid Bill Shock

This is the sticky gotcha from the Luna guide and the Sol vs Terra vs Luna comparison — short enough to bookmark, expensive enough to ignore.

What the alias actually does

OpenAI’s short family alias gpt-5.6 resolves to the Sol tier — the expensive, flagship member of the family. It does not mean “whatever GPT-5.6 is cheapest this week,” and it does not mean Luna.

That naming pattern matches how many vendors treat a bare version string as the flagship. The trap is that Luna’s marketing is loud after the −80% cut, so builders type gpt-5.6 and assume they got the bargain model.

How expensive the mistake is

Rough list-rate ratio after July 30, 2026 (per 1M tokens, input/output):

• Luna gpt-5.6-luna — $0.20 / $1.20
• Terra gpt-5.6-terra — $2 / $12 (~10× Luna)
• Sol gpt-5.6 or gpt-5.6-sol — ~$5 / $30 (~25× Luna input)

On agent workloads (tool loops, long schedules), a silent Sol route is not a “slightly higher” bill — it is a different budget class. Confirm the model field in usage logs, not only the string you thought you configured.

How to fix it

Pin the slug you mean in every client, gateway, and fallback:

• Set model: "gpt-5.6-luna" (or terra / sol) explicitly — never rely on gpt-5.6 for Luna
• Search configs and IaC for bare gpt-5.6 and replace
• On Chat Completions with tools, Luna still needs reasoning_effort: "none" (or use Responses) — separate from the alias issue, but the same “it 400s after I switched models” debugging trail
• Re-check OpenRouter / proxy aliases: third-party names do not always mirror OpenAI’s routing

CloudAxis note

Operators on CloudAxis do not pick model strings in a dropdown. Auto Mode is powered by GPT-5.6 Luna on every plan including Free. This post is for API builders and anyone migrating stacks after the July 30 price cut.

Frequently asked questions

Does gpt-5.6 mean Luna?

No. The gpt-5.6 alias routes to Sol (flagship). Use gpt-5.6-luna for the cheap tier.

What model ID should I use for the cheapest GPT-5.6?

gpt-5.6-luna — $0.20/$1.20 per 1M tokens on the July 30, 2026 OpenAI list.

Is gpt-5.6-sol different from gpt-5.6?

They should resolve to the same Sol tier. Prefer the explicit gpt-5.6-sol id in configs so intent is obvious in code review.

How do I verify which model ran?

Check the model field on usage events / completion metadata from OpenAI (or your proxy). Do not trust the config filename alone.