GPT-6 Sol/Luna Begins Competing on Cost-Effectiveness: Sol Costs Only Half the Price of Opus 5.5

Beating AI News Flash: OpenAI releases GPT-6 Sol and GPT-6 Luna, bringing GPT-6 capabilities down to two cheaper tiers of models. Sol focuses on complex coding and professional work, while Luna targets high-frequency everyday tasks. API pricing per million tokens is $2 input and $10 output for Sol, and $0.10 input and $0.50 output for Luna; this is at least 50% cheaper than the promotional price of GPT-5.6. Claude Opus 5.5, released the same day, is priced at $4 and $20, meaning Sol's base token unit price is exactly half.

In OpenAI's published AutomationBench, Sol with xhigh effort enabled scored 33.2%, higher than Opus 5's 26.9%, with a cost per question of only $0.27, about 9% of the latter. The comparison here is with Opus 5, not the newly released Opus 5.5. On DeepSWE v1.1, Sol with max enabled scored 68.8%, only 1.1 percentage points below Fable 5's highest score, but with a cost per question about 80% lower.

Safety evaluations also showed notable changes. In tests deliberately designed to induce models to misreport coding work, Sol's deception rate dropped from 10.4% in GPT-5.6 Sol to 1.3%, and Luna's fell from 9.5% to 2.8%; GPT-6 Astra was 0.5%.

The two models have already entered the API, ChatGPT Work, and Codex. Free and Go users can use Luna on desktop, but it has not yet entered Chat. GPT-6's cached reads also enjoy a 90% discount, and adjusting reasoning effort or toggling tools will not break previous cache.