OpenAI has launched GPT-6 Sol and GPT-6 Luna, pushing its mainstream AI models forward less than three months after GPT-5.6 arrived. This time, however, the most important part of the upgrade may not be raw intelligence. It is how much cheaper OpenAI says that intelligence has become.
The company claims GPT-6 Sol makes roughly half as many mistakes as GPT-5.6 Sol, while the lower-cost GPT-6 Luna can reach performance comparable to the previous-generation Sol at higher reasoning settings. Those are OpenAI’s own measurements, so independent testing will be needed to establish how consistently the improvements translate into everyday workloads.
Pricing is arguably the bigger story. OpenAI says API prices for Sol and Luna have been cut by 50% compared with GPT-5.6 promotional pricing. That matters particularly for developers running agents, coding systems and other applications where millions of tokens can quickly turn modest price differences into substantial infrastructure bills.
OpenAI’s benchmark results reinforce that cost-first strategy. The company says GPT-6 Luna improved over GPT-5.6 on its professional-work benchmark while costing 58% less per task. GPT-6 Sol is also being positioned against Anthropic’s Claude family, with OpenAI reporting competitive results across professional workflows, coding and software engineering at substantially lower costs in several tests.
As always, vendor benchmarks deserve caution. Model performance can change considerably depending on prompts, reasoning settings and workloads, while comparisons against rival models can become outdated almost immediately as competitors release updates.
GPT-6 also introduces improvements beyond benchmark scores. OpenAI says Sol and Luna should produce clearer, slightly shorter responses with less jargon and unnecessary detail. Prompt caching has also been improved, with the company claiming cached prompts can cost up to 90% less to process than uncached ones — potentially useful for agents repeatedly working with the same instructions and context.
Safety results remain more complicated. OpenAI reports improvements in several tests involving unauthorized actions, including a simulated scenario where GPT-6 Sol followed prohibited instructions in 11% of applicable cases, compared with 52% for GPT-5.6 Sol. That is a substantial reduction, but hardly a reason to abandon human oversight for sensitive autonomous workflows.
GPT-6 Sol and Luna are available now. The short gap since GPT-5.6 makes the release notable, but the broader shift is clearer: the AI race is increasingly about delivering capable models economically enough to run them everywhere.
