GPT-6 Sol and Luna: half price, and permanent this time

OpenAI released GPT-6 Sol and GPT-6 Luna today, and the GPT-6 pricing is the whole story. Sol lands at $2 per million input tokens and $10 per million output. Luna lands at $0.10 and $0.50. Those are roughly half what the GPT-5.6 versions cost under a promotion that was set to lapse in November — and OpenAI says these are list prices, not a countdown.

What actually happened

OpenAI introduced Sol and Luna as the mid-tier and budget members of the GPT-6 family, under the flagship Astra that shipped earlier this month. They are trained with the same methods as Astra, and OpenAI credits caching and inference improvements for the lower serving cost.

The deltas, against the GPT-5.6 models they replace:

• Sol: $4/$20 → $2/$10 per million tokens. 50% off both sides.
• Luna: $0.20/$1.20 → $0.10/$0.50. 50% off input, 58% off output.

An OpenAI spokesperson told VentureBeat these are "permanent prices, not promotional or introductory pricing." That sentence is the news. We wrote in August that the $4/$20 rate on GPT-5.6 Sol was a promotion with a November 21 end date and told you to price for the reversion. The reversion is now moot — the successor model undercuts the promo permanently.

Both are live in the API as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. Free and Go users get Luna in the desktop app.

Why permanent AI pricing matters for your business

A promotional rate and a list rate are different inputs to the same spreadsheet, and only one of them survives contact with a customer contract. If you have been quoting per-document or per-ticket pricing built on a discount with a calendar date, you can now rebuild it on a number that is not scheduled to move against you.

Notice what happened today across two vendors. Anthropic cut Opus 5.5 to $4/$20 this morning. OpenAI put Sol at $2/$10 hours later. Frontier-adjacent inference is repricing on a weekly clock, not a yearly one. Loyalty to whoever is cheapest this Tuesday is not a strategy — a thin abstraction layer is.

Three moves worth making before the week ends:

Re-run the tier question, not just the price. Luna at $0.10/$0.50 is cheap enough that work you kept on a mid-tier model for cost reasons may belong on the budget tier now. Classification, routing, extraction, tagging — test them on Luna before you assume you need Sol.

Pin the model string on purpose. gpt-6-sol is not a drop-in for gpt-5-6-sol. Same family, different behavior. Pin deliberately, then schedule the review instead of discovering the change in a customer-facing output.

Re-run your evals before you migrate. OpenAI's benchmarks describe OpenAI's workload. Twenty of your own test cases and an afternoon tells you more than any leaderboard.

Sources: OpenAI, VentureBeat.