Skip to content

AI

GPT-6 Sol and Luna cut prices in half. Your salary stayed put.

OpenAI's updated Sol and Luna land at roughly half the old API price, with intelligence scores that stay level. Cheaper robot brains, same robot brains.

Cover photo: HaeB / CC BY-SA 4.0

TL;DR

  1. Sol now costs $2/$10 per million tokens; Luna costs $0.10/$0.50. Both are about half the old price.
  2. Artificial Analysis says benchmark scores stay level with GPT-5.6: some gains, some regressions.
  3. OpenAI credits caching and inference improvements for the price cut, not a smarter model.

What happened

Weeks after launching GPT-6 Astra, which OpenAI pitched as its most capable model yet, the company has refreshed the smaller tiers: GPT-6 Sol and GPT-6 Luna.

Sol is built for complex work like coding. Luna is meant for high-volume clerical jobs, such as summarizing documents, extracting information and answering quick questions.

The headline is the API bill. Sol falls from $4/$20 to $2/$10 per million input/output tokens. Luna falls from $0.20/$1.20 to $0.10/$0.50. OpenAI attributes the drop to improvements in caching and inference.

Independent benchmarker Artificial Analysis ran the numbers. Luna comes in at $0.07 per task, against $0.18 for its predecessor, roughly 60% less. Sol's cut is roughly 50%.

Here is the catch: the savings come from the price list, not from efficiency magic. Both new models actually use slightly more output tokens per task, 31k versus 29k for Sol and 51k versus 41k for Luna.

Why it matters

This is a price story, not a brains story. Artificial Analysis says Intelligence Index and Coding Agent Index scores stay level with GPT-5.6, with progress on some evaluations and regressions on others.

In OpenAI's Codex harness, Sol scores 57 on the Coding Agent Index, up 2 points. Luna, meanwhile, regresses on that same coding test. So Luna is cheaper and a bit worse at coding, which is fine if you wanted it for clerical work anyway.

Same intelligence, half the invoice: that is the pitch, and the benchmarks mostly back it up.

For developers, the math changes quickly. Workloads that were borderline at the old price, like bulk document processing or agents that burn through tokens, get easier to justify.

Cache reads keep their 90% discount and cache writes keep their 25% premium, so the usual optimization tricks still apply.

For the rest of us, the trend line is the point. Each generation, the same level of machine competence gets cheaper. The humans doing the clerical work Luna was named for may want to read that twice.

What to do with it

  • Re-run your cost model.Plug the new $2/$10 and $0.10/$0.50 rates into your estimates before deciding a feature is too expensive to ship.
  • Test Luna on the boring stuff.Summaries, extraction and quick Q&A are exactly what OpenAI positions it for, so benchmark it there first.
  • Do not assume a coding upgrade.Sol improved on the coding index but Luna slipped, so check your own coding workloads before swapping models.
  • Keep your caching tight.The 90% cache-read discount still stacks with the lower base price, so reuse prompts wherever you can.

Ciao, sapiens. The robots got a discount; go check whether your budget did too.