Three weeks after the launch of the GPT-5.6 family, OpenAI has significantly reduced the prices for two of its models. Starting July 30, 2026, Luna will cost 80 percent less via the API, and Terra will cost 20 percent less. The more expensive model Sol remains unchanged in price but will receive a new, faster processing mode for enterprise customers.
Luna now costs only one-fifth of the previous price
For GPT-5.6 Luna, the cheapest model of the family launched in July, the price per million input tokens drops from one US dollar to 20 US cents, and the output price from six to 1.20 US dollars. In everyday terms: 20 US cents will now cover about 750,000 words of input text with Luna, whereas the same amount previously cost a full US dollar. Terra, the middle tier, will be 20 percent cheaper and now costs 2.00 US dollars per million input tokens and 12.00 US dollars per million output tokens. The more expensive flagship model Sol retains its price of five and 30 US dollars per million tokens, as OpenAI states in a blog post. Instead, Sol will receive a new Fast mode that processes requests for enterprise customers up to 2.5 times faster and replaces the previous Priority Processing option; ongoing requests marked as “priority” will continue to function unchanged. OpenAI had introduced the GPT-5.6 family only on July 9, 2026, along with the agent ChatGPT Work, at the now-reduced starting prices.
OpenAI points to efficiency gains in model delivery
As a reason for the price cut, the company cites advancements in serving and inference, the ongoing operation of the models after training is completed. Part of this is attributed to GPT-5.6 Sol itself: the model independently rewrote production code for the graphics processors used, thereby reducing delivery costs by 20 percent and improving overall efficiency by more than 15 percent. Additionally, there are optimizations across the entire software and hardware stack, including improved hardware routing, more efficient inference software, and smarter context caching. OpenAI promotes the move with the formula “more intelligence per dollar” and positions the price reduction as a direct result of technical improvements, not merely a reaction to market pressure.
The fact that an AI model like Sol contributes to optimizing its own operating costs underscores, according to OpenAI, the practical utility of the systems beyond mere text generation. How robust the stated efficiency values are cannot be verified externally; independent measurements of the actual serving costs of large language models have hardly been published in the industry.
The move clearly targets cheaper competing models
The new prices clearly target providers with cheaper inference, including Google’s Gemini 3.5 Flash-Lite and several Chinese providers, as the trade publication eWeek categorizes it. OpenAI itself is positioning Luna aggressively: according to its own calculation, as reported by eWeek, Luna is about 99 percent cheaper in cost comparison per completed task than Anthropic’s model Fable 5 – a company-specific figure that is not independently verified. For companies, the price cut is intended to make multi-stage agent workflows such as code generation, document processing, or automated request routing more economically viable. OpenAI also promotes so-called mixed model architectures, where companies route routine tasks to the cheaper models Luna or Terra and reserve only complex tasks for the more expensive Sol. Competition among AI providers is thus increasingly shifting from pure performance comparisons to cost accounting per completed task.
It remains to be seen whether Anthropic or Google will respond to the price cut with their own price reductions or whether they will focus on higher model quality instead of price. For companies deploying AI agents at scale, this cost question will determine which providers remain viable for routine tasks such as automated document processing going forward.
