OpenAI Adds GPT-6 Sol and Luna at Half the Price of GPT-5.6
OpenAI introduced GPT-6 Sol and GPT-6 Luna as additions to the GPT-6 model family. In a Sept. 16 post, the company identified the two models as new members of that family. OpenAI
OpenAI said it trained Sol and Luna with similar methods as GPT-6 Astra. API prices, the rates developers pay to send text in and receive text back, for Sol and Luna are 50% lower than their GPT-5.6 promotional pricing.
For Sol, OpenAI claimed improved factual reliability. The company said Sol makes about half as many mistakes as its predecessor on its internal factuality evaluation based on real-world conversations where users flagged mistakes.
On agentic work, where a model completes multi-step tasks using software tools, OpenAI pointed to AutomationBench. The company said GPT-6 Sol at xhigh effort, a high reasoning setting, outperforms Claude Opus 5 at max effort on AutomationBench at 9% of Opus 5's cost per task.
The pricing comparison refers back to the GPT-5.6 cycle. GPT-5.6 shipped in three variants: Sol, Terra and Luna, with Sol as the workhorse, Terra as the intermediate option, and Luna as the budget-friendly option. TechCrunch Initial costs were $5 per million input tokens and $30 per million output tokens for Sol, and $1 per million input tokens and $6 per million output tokens for Luna. Tokens are the small chunks of text models process. TechCrunch
OpenAI then discounted that family. It dropped API and credit pricing for GPT-5.6 Sol by over 20% for three months and cut GPT-5.6 Luna by 80% as its most affordable model. OpenAI After that cut, Luna was listed at $0.20 per million input tokens and $1.20 per million output tokens. Yahoo Finance
That rollout did not start as open access. OpenAI had limited GPT-5.6 access to vetted partners before broader availability, with Reuters reporting the company said Sol, Terra and Luna would launch on Thursday. Reuters Early descriptions framed Sol as a next-generation model with stronger capabilities in coding, science and cybersecurity, and Luna as a fast and affordable model. OpenAI
The broader context here is how quickly price becomes architecture. When input and output token rates fall by half against an already discounted baseline, teams can afford longer contexts, the amount of text a model can consider at once, plus more tool calls and higher reasoning effort in production. That factuality evaluation weights errors users actually notice and report in dialogue, not curated academic sets. The AutomationBench claim couples two variables engineers track separately, reasoning effort and cost per completed task.
In my view, the two claims to test first are factuality and cost per task. Halving user-flagged mistakes, if it holds across domains, lowers review overhead and exception handling. That is often the real blocker for deployment. And an AutomationBench win at 9% of the cost per task reframes build-versus-buy for automation pipelines. Engineers will want to reproduce both under their own traces, with their own failure taxonomies and retry budgets. Worth flagging, internal evaluations and vendor-run cost math are starting points, not substitutes for that work.
Looking at what this means for deployment, the practical effect is straightforward. More reasoning per dollar, with fewer corrections downstream. That combination is what makes model updates stick in production systems.
