OpenAI has started previewing its GPT-5.6 series, which will be available in three versions, to a limited number of trusted partners. The company says the Sol variant is its most robust model yet, while the Terra is for everyday use and performs similarly to the GPT-5.5 despite being twice as cheap. The Luna, the last variant, is the company’s least expensive model. OpenAI plans to give them a wide release sometime in the coming weeks.
The company gave the US government a preview of GPT-5.6 and its capabilities earlier today. It is also a request for management to first test the model on a small group of trusted partners in their “shared partnership” with the government. “We do not believe this type of government access process should be the default for the long term,” OpenAI wrote in its announcement. It said it was taking a “temporary step,” for now, as it confirmed it could release its latest model series to the public soon.
President Trump signed an AI cybersecurity order earlier this month, asking companies to submit their powerful models for the government’s voluntary review 30 days before they become publicly available. According to a recent report by The New York TimesOpenAI, Anthropic, Google, xAI and Microsoft have been giving the government early access to their latest models even before Trump signed the order. Meta was the only participant, and the US government was reportedly urging it to submit its AI models for testing.
GPT-5.6 introduces a “big” thinking effort, giving Sol more time to think critically. Sol is also OpenAI’s most powerful cybersecurity model and is the best choice to help users detect and fix vulnerabilities. OpenAI says Sol comes with strong protection for high-risk operations and sensitive applications. It also says that the company has spent several weeks finding its weaknesses and strengthening them against real-world attacks.
The company puts safeguards on all exceptions, however, to ensure that it holds up to real pressure from adversaries. In addition, OpenAI trained GPT-5.6 to reject “forbidden online help,” including attempts to jailbreak the model. It has spent 700,000 GPU hours finding jailbreaks around the world to develop countermeasures, and promises to use “a rapid response process to reproduce, evaluate, prioritize, and remediate newly discovered jailbreaks.”
OpenAI’s focus on avoiding jailbreaks may have come from the Anthropic experience. A few weeks ago, Anthropic stopped all access to its Mythos 5 and Fable 5 models after an order from the government. Although the company did not say specifically, Amazon and other companies have reportedly notified authorities that their models can be seized and used for malicious purposes. It has begun to raise its reach, however, as the US government recently granted Anthropic permission to redistribute Mythos to a select group of organizations.
The company has priced GPT-5.6 Sol at $5 per million input tokens and $30 per million output, which is significantly less than what Fable was worth when it was still available. ($10 to input and $50 to output for the same amount of tokens.) Terra costs $2.50 to input and $15 to output, while Luna costs $1 to input and $6 to output.
