OpenAI halves Sol and Luna to $2/$10 and $0.10/$0.50, and independent testing finds their intelligence level with GPT-5.6 (2026)

OpenAI put two models below its flagship on September 22 and priced both at half what the GPT-5.6 models they replace cost on promotion. GPT-6 Sol now runs $2 per million input tokens and $10 per million output; GPT-6 Luna runs $0.10 and $0.50. 1 Both were trained with methods similar to GPT-6 Astra, the September flagship that OpenAI still calls its best model across the board. 1

The two are live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu accounts, with Free and Go users reaching Luna in the desktop app. The API serves them as gpt-6-sol and gpt-6-luna, and ChatGPT's Chat surface gets them later, with the rollout going out through the day. 1

Signal	Confirmed detail	What it means for you
What shipped	Two tiers under Astra, trained with Astra's methods, Sol above Luna. 1	Astra stays the pick for the hardest work.
What it costs	Sol $2/$10, Luna $0.10/$0.50 per million input/output tokens, with cached input reads at a 90% discount. 1	Cache-heavy agent loops take the largest share.
Where the intelligence lands	Artificial Analysis measures Sol at 48 and Luna at 37 on its Intelligence Index, against GPT-5.6 Sol's 47 and Astra's 53. 2 3 4 5	The cut buys the same measured level as the model it replaces.
What gives	Sol drops about 100 Elo on GDPval-AA v2.1, Luna about 75, and Luna a further 45 on AA-Briefcase v1.1. 6	Long knowledge-work deliverables regressed.

Where OpenAI's own numbers stop

OpenAI's reliability headline rests on an internal evaluation built from de-identified conversations where users had already flagged a mistake. OpenAI notes those conversations are unrepresentative of typical use, and the scores are uncontrolled for answer length. 1 Its AutomationBench table prices Claude Fable 5.1 without the Opus 5 fallbacks that ran on about 40% of tasks; OpenAI says leaving them out understates that row's cost. 1 The computer-use row shows the pattern most plainly: Sol at its highest effort setting scores 60.5% where Claude Opus 5 at medium effort scores 60.3%. 1 ZDNET read that as OpenAI's top setting matching an older Anthropic model at half setting. 7

Where Artificial Analysis puts them

On its Intelligence Index, the evaluator finds Sol level with GPT-5.6 Sol and Luna level with GPT-5.6 Luna, with Sol five points under Astra, a smaller step than the price cut suggests. 6 The evaluator's cost per task falls with the price: $1.06 for Sol against GPT-5.6 Sol's $1.99, and $0.07 for Luna against $0.18. It attributes that to the price cut, since both models spend slightly more output tokens per task than the versions they replace. 6 Luna loses 2 points on the Coding Agent Index, which runs inside OpenAI's Codex harness, while Sol gains 2. 6

Two stacked Artificial Analysis panels: a bar chart of Intelligence Index scores showing GPT-6 Sol at 48, GPT-5.6 Sol at 47 and GPT-6 Astra at 53, and a scatter plot of index score against cost per index task with GPT-6 Sol and GPT-6 Luna on the Pareto line — Artificial Analysis's Intelligence Index, and the same index plotted against cost per index task, in the figure the evaluator published with its measurements; the lower panel puts GPT-6 Sol and GPT-6 Luna on the cost frontier. 6

The reliability claim looks different on the evaluator's knowledge benchmark. Sol cuts its hallucination rate from 92% to 60%, and its accuracy falls as well, from 59% to 54%, because it now answers 83% of the questions where GPT-5.6 Sol answered 99%; Luna's accuracy holds at 44%. 6

Why it matters

Anthropic released Claude Opus 5.5 within hours of this launch, and Fortune reports both companies now competing on cost per task as businesses audit their AI spending. 8 Reuters reads the release as offering some of Astra's capabilities at half the promotional rates of its predecessor. 9 For ChatGPT subscribers the cut lands on the API, and ZDNET notes the announcement leaves open whether plan limits stretch further. 7