OpenAI cuts GPT‑6 prices in half. Sol aims to bring frontier AI into everyday business work

OpenAI is expanding the GPT‑6 family with the Sol and Luna models while opening another round of price competition among makers of the most advanced AI. GPT‑6 Astra remains the company’s most capable model, but Sol and Luna are intended to bring some of its capabilities into much cheaper operation. GPT‑6 Sol input pricing falls from $4 to $2 per million tokens compared with GPT‑5.6 Sol, and output pricing from $20 to $10. Luna is priced at $0.10 for input and $0.50 for output. For companies, the important point is not merely that one query costs less. Agent systems can make dozens or hundreds of model calls during a single work task, use tools, check their own results and delegate parts of the work to other agents. A difference of several dollars per million tokens therefore multiplies at scale. OpenAI says Sol has improved over the previous generation in professional work, programming, computer use and factual accuracy. On AutomationBench it scored 33.2 percent at the highest setting, with an average cost of $0.27 per task. The company says Claude Opus 5 scored 26.9 percent at more than eleven times the cost. These are, however, a benchmark and configuration presented by OpenAI; a company should base its decision on its own work tasks. The improvement in the cheaper Luna is notable. OpenAI says that at higher reasoning effort it reaches the level of the previous Sol in some tasks at roughly one hundredth of the cost. Economically, that may matter more than a few extra percentage points from the most expensive model. Companies do not need to use the most capable model for everything. Simple document classification, data checks or routine communication can be handled by a cheaper model, while more complex analysis is routed to Sol or Astra. Intelligent routing of tasks among models is therefore becoming a separate way to save money. Factual reliability is also improving. OpenAI says Sol makes roughly half as many errors as its predecessor in an internal test based on real user reports. Even that figure does not mean that result checking can be switched off. In financial, legal or operational work, what matters is the absolute frequency of serious errors, not only relative improvement. The pricing trend has a broader business impact. If model capabilities rise at the same time as unit prices fall sharply, the return-on-investment threshold for automation drops. A process that was too expensive a year ago for continuous AI use may now make economic sense. Management should therefore not evaluate every new model generation as a separate software purchase. It is better to recalculate the cost per correctly completed work task continuously. That is the measure that shows whether model improvement and lower prices actually translate into productivity. KEY TERMS Token: A basic unit of text consumption commonly used to price language-model operation. Cost per task: The total model cost required to complete a specific work activity. Model routing: Automatic selection of a cheaper or more capable model according to task complexity.

OpenAI cuts GPT‑6 prices in half. Sol aims to bring frontier AI into everyday business work

The full article is available to users registered at Hard Skills