en.Wedoany.com Reported - OpenAI has initiated limited testing of the GPT-5.6 series with select trusted partners, marking a new phase in the AI industry. The company has introduced three model variants with different capabilities and prices, a decision closely tied to U.S. government requirements. Prior to public release, OpenAI previewed the model to the U.S. government, responding to President Trump's AI cybersecurity executive order, which requires companies to voluntarily submit their most powerful models for government review 30 days before public release. OpenAI emphasized that this government access process should not become a long-term standard, but only a short-term step to ensure models are made available to the public as quickly as possible.
The GPT-5.6 series includes three variants, each with distinct specifications and target uses. The Sol variant is considered OpenAI's most powerful model to date, designed for complex tasks in the cybersecurity domain, helping users discover and fix system vulnerabilities. This model introduces a "maximum reasoning effort" feature, providing deeper thinking time, with OpenAI spending weeks identifying its weaknesses and reinforcing it against real-world attacks. The Terra variant targets everyday use, with performance similar to GPT-5.5 but at half the cost, offering users a cost-effective option. The Luna variant is the lowest-cost among the three, suitable for lighter tasks.
In terms of pricing, OpenAI has set competitive rates for the three variants. The Sol variant is priced at $5 per million input tokens and $30 per million output tokens, lower than Anthropic's Fable model ($10 per million input tokens and $50 per million output tokens). The Terra variant is priced at $2.50 per million input tokens and $15 per million output tokens. The Luna variant is the most affordable, at $1 per million input tokens and $6 per million output tokens. OpenAI plans to broadly release these three variants in the coming weeks.
GPT-5.6 has invested significant effort in safety, with all variants implementing enhanced protections for high-risk activities and sensitive requests. OpenAI trained the model to refuse "prohibited cyber assistance," including attempts at jailbreaking. To ensure safety, OpenAI spent 700,000 GPU hours identifying universal jailbreak methods and developing preventive measures, while committing to a "rapid response process" to address newly discovered jailbreak vulnerabilities. The emphasis on jailbreak prevention likely stems from recent incidents at Anthropic, where its Mythos 5 and Fable 5 models were suspended due to potential jailbreaking for malicious purposes.
This limited testing initiative has broad implications for the AI industry. Government regulation is beginning to influence how companies release models, with Trump's cybersecurity executive order setting a precedent for government approval before releasing the most powerful models. Meanwhile, price competition between OpenAI and Anthropic is intensifying, with Sol's pricing below Fable indicating an intent to attract enterprise customers. Additionally, the focus on jailbreak prevention suggests AI model safety has become an industry priority. According to The New York Times, OpenAI, Anthropic, Google, xAI, and Microsoft provided early model access to the government before the executive order was signed, with Meta being the only company to refuse, continuing to face government pressure to submit models for evaluation.
OpenAI's limited release of GPT-5.6 at the Trump administration's request highlights the increasingly complex dynamic between technological innovation and government regulation. Meanwhile, the development of the custom AI chip Jalapeño demonstrates OpenAI's commitment to reducing reliance on Nvidia.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
