GPT-6 Sol and Luna: half the price, same intelligence, and that is the point
OpenAI dropped GPT-6 Sol and GPT-6 Luna on September 22, nineteen days after Astra. Sol is $2 in and $10 out per million tokens. Luna is $0.10 and $0.50. Both are exactly half of what the GPT-5.6 versions cost under promotional pricing. Independent analysis says raw intelligence sits at GPT-5.6 levels, up on some evals and down on others.
That reads like a disappointing release until you look at what these models are actually for. On AutomationBench, Sol at xhigh scores 33.2% at $0.27 per task, beating Claude Opus 5 at max effort (26.9%) while costing about a ninth as much. Luna gained 5.4 points while costing 58% less than its predecessor. On DeepSWE v1.1 Sol reaches 68.8% at $1.00 per task, and Luna matches that same 68.8% for $0.22. A model a twentieth the price of the flagship hitting the same agentic coding score is a pricing event, not a capability event.
The honest caveat is that Sol's ceiling is lower than what came before it. Its best DeepSWE (68.8%) and OSWorld 2.0 (64.4%) scores are below GPT-5.6 Sol's best (72.7% and 66.2%) and below Claude Opus 5's (73.7% and 70.2%). Astra stays the performance model. Sol and Luna are the models you run a hundred thousand times.
That is the split worth paying attention to. For most of the last two years the interesting question was which model is smartest. For agent workloads the question has quietly become which model gives you the most successful task completions per dollar, and those two questions have different answers. Both OpenAI and Anthropic cut prices on the same day this week. Neither cut is charity.
Available now in ChatGPT Work and via the API as gpt-6-sol and gpt-6-luna, with ChatGPT rollout staged.
https://openai.com/index/introducing-gpt-6-sol-and-luna/
← Back to all articles
That reads like a disappointing release until you look at what these models are actually for. On AutomationBench, Sol at xhigh scores 33.2% at $0.27 per task, beating Claude Opus 5 at max effort (26.9%) while costing about a ninth as much. Luna gained 5.4 points while costing 58% less than its predecessor. On DeepSWE v1.1 Sol reaches 68.8% at $1.00 per task, and Luna matches that same 68.8% for $0.22. A model a twentieth the price of the flagship hitting the same agentic coding score is a pricing event, not a capability event.
The honest caveat is that Sol's ceiling is lower than what came before it. Its best DeepSWE (68.8%) and OSWorld 2.0 (64.4%) scores are below GPT-5.6 Sol's best (72.7% and 66.2%) and below Claude Opus 5's (73.7% and 70.2%). Astra stays the performance model. Sol and Luna are the models you run a hundred thousand times.
That is the split worth paying attention to. For most of the last two years the interesting question was which model is smartest. For agent workloads the question has quietly become which model gives you the most successful task completions per dollar, and those two questions have different answers. Both OpenAI and Anthropic cut prices on the same day this week. Neither cut is charity.
Available now in ChatGPT Work and via the API as gpt-6-sol and gpt-6-luna, with ChatGPT rollout staged.
https://openai.com/index/introducing-gpt-6-sol-and-luna/
Comments