The AI race usually focuses on which company has the smartest model, but intelligence is only part of the equation. If using that model costs a fortune, plenty of developers and businesses simply cannot justify throwing it at every task.
OpenAI appears to understand that problem. The company has introduced GPT-6 Sol and GPT-6 Luna, two new models that bring technology developed for GPT-6 Astra to cheaper tiers. More importantly, OpenAI is cutting API prices substantially at the same time.
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna drops all the way to $0.10 per million input tokens and $0.50 per million output tokens. OpenAI says those prices represent a 50 percent reduction compared with the promotional pricing of their GPT-5.6 equivalents.
That could matter more than another small benchmark victory. AI agents can burn through enormous amounts of tokens as they browse files, write code, use tools, reconsider decisions, and work through long conversations. Cutting token prices gives developers much more freedom to let those systems work without constantly worrying about the meter running.
OpenAI is also claiming some impressive performance for the money. On AutomationBench, which measures AI agents performing business workflows across areas such as sales, marketing, finance, support, and HR, GPT-6 Sol at xhigh reasoning effort scored 33.2 percent. Claude Opus 5 at maximum effort scored 26.9 percent in OpenAI’s comparison, while costing 11.1 times as much per task.
The coding numbers tell a similar story. On DeepSWE 1.1, GPT-6 Sol scored 68.8 percent at maximum effort. Claude Fable 5 reached 69.9 percent at xhigh effort, but OpenAI says Sol accomplished its result at roughly 80 percent lower cost per task. GPT-6 Luna scored 66.6 percent, with OpenAI claiming substantially lower costs than the Claude models used in the comparison.
Of course, these are benchmark results presented by OpenAI, so they should not be confused with a guarantee about how every model will perform in the real world. Developers care about benchmark scores, but they also care about latency, reliability, tool use, context handling, and how often an agent screws something up after spending several minutes working on a task.
Factual accuracy is another area OpenAI says it has improved. In an internal evaluation based on de-identified ChatGPT conversations where users had previously flagged factual mistakes, GPT-6 Sol made about half as many errors as GPT-5.6 Sol. OpenAI notes that these conversations were deliberately selected because they had produced errors before, so the results are not representative of normal ChatGPT usage.
The new models should apparently be less annoying to talk to as well. OpenAI says GPT-6 Sol and Luna use less jargon and fewer unnecessary details, while generally providing slightly shorter answers without sacrificing substance. Anyone who has watched an AI model spend six paragraphs answering something that required three sentences may appreciate that particular improvement.
There is another potentially important change hiding behind the lower token prices. OpenAI has improved prompt caching for GPT-6, allowing developers to receive a 90 percent discount on cached input-token reads. Developers can also change reasoning effort or available tools without automatically losing previously cached context.
That could make long-running agents considerably cheaper. GitHub says improvements to OpenAI’s caching system have reduced the share of prompt tokens requiring fresh processing by more than 50 percent across billions of requests, according to OpenAI. For applications repeatedly feeding the same instructions, documents, or conversation history into a model, those savings can add up quickly.
GPT-6 Astra still sits at the top of OpenAI’s model family when maximum performance matters, but Sol may become the more interesting option for many developers. Getting reasonably close to flagship performance while spending considerably less money is often a better trade than chasing the last few percentage points on a benchmark.
Availability is somewhat complicated at launch. GPT-6 Sol and Luna are available starting today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers. Free and Go users can access GPT-6 Luna through the desktop app. OpenAI says the models are not yet generally available in Chat, with access rolling out gradually. Developers can already use them through the API as gpt-6-sol and gpt-6-luna.
AI companies will undoubtedly continue fighting over benchmark charts, but the more consequential battle may increasingly be over economics. A model does not need to be the absolute smartest available if it delivers nearly the same result for a fraction of the cost.
With GPT-6 Sol and Luna, OpenAI is betting that cheaper intelligence can be just as important as smarter intelligence. If the company’s real-world performance matches its benchmark claims, that could make GPT-6 useful in far more places than another expensive flagship model ever would.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz
