OpenAI has released two newcomers to its GPT-6 family of models, promising lower token costs, more intuitive communication, and an ongoing effort to prevent the kind of rogue agent behavior that has hung heavy over the company in recent months and added more fuel to already heated AI safety debates.
GPT-6 Sol and Luna, launched Tuesday afternoon, cost about half as much as their GPT-5.6 counterparts, which made their public debut in July following a staggered release overseen by the federal government. They’re being marketed as lighter-weight and more affordable versions of GPT-6 Astra, the blockbuster model released earlier this month, which OpenAI president Greg Brockman reportedly described as the first harbinger of “the AGI era.”
OpenAI said in its announcement on Tuesday that GPT-6 Sol and Luna were trained using “similar methods as GPT-6 Astra,” which could include recurrent depth, a controversial technique that can enhance a model’s performance but make it harder for human researchers’ ability to monitor its decision-making processes. The company didn’t immediately respond to a request for comment about the use of recurrent depth to train the two new models.
“Together, these improvements make advanced AI practical for more everyday tasks and applications at scale,” OpenAI wrote in its announcement.
Efficiency and affordability have become a major focus for OpenAI and its competitors at a time when growing numbers of their customers, including both individuals and businesses, are complaining about the steep and sometimes unpredictable cost of token usage. Anthropic also released a new model on Tuesday, called Claude Opus 5.5, which the company says costs 40% less for typical workloads compared to its predecessor. Both OpenAI and Anthropic also emphasized improvements in communication in their respective model announcements on Tuesday: “Expect to see more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall without losing substance,” OpenAI wrote of GPT-6 Sol and Luna.
The company added that its models have been trained with the same alignment safeguards as GPT-6 Astra, which OpenAI describes as its “most aligned model to date.” But it also acknowledges that alignment is a young science and that much remains mysterious, as the July Hugging Face hack—which involved thousands of agents conspiring with each other via a makeshift messageboard—made disconcertingly clear. In the GPT-6 Astra system card (updated today with an appendix about GPT-6 Sol and Luna), the company said it’s continuing to study “emerging patterns of agent-to-agent communication, including how independently tasked agents discover one another, exchange information, and coordinate their actions.”
Read the full article here
