OpenAI has begun previewing its new GPT-5.6 series, limited to a small group of trusted partners, after the US administration requested a voluntary review of its most powerful models under President Trump's cybersecurity executive order signed earlier this month. The GPT-5.6 family includes three variants: Sol, Terra, and Luna, each with distinct capabilities and pricing.
Sol is OpenAI's most powerful model with deep reasoning capabilities
Sol represents the pinnacle of OpenAI's current capabilities. The company states that Sol can dedicate more time to reasoning via a new 'max reasoning effort' parameter, allowing it to deeply analyze complex problems. It is also the best model for cybersecurity, designed to help users find and fix vulnerabilities. OpenAI invested several weeks identifying Sol's weaknesses and fortifying it against real-world attacks. Pricing is set at $5 per million input tokens and $30 per million output tokens, significantly cheaper than Anthropic's Fable model which cost $10 and $50 respectively.
Sponsored Protocol
Terra offers GPT-5.5-like performance at half the price for everyday use
Terra is designed for daily tasks, with performance comparable to GPT-5.5 but at half the cost: $2.50 for input and $15 for output per million tokens. This variant aims to make advanced AI more accessible without compromising quality. OpenAI applied safeguards across all variants to withstand real adversarial pressure, including jailbreak attempts.
Luna is the most affordable variant with costs reduced by 90% compared to top-tier models
Luna completes the trio as an ultra-low-cost model: $1 for input and $6 for output per million tokens. It is intended for lightweight applications and developers needing economical, scalable AI. OpenAI trained GPT-5.6 to refuse 'prohibited cyber assistance' and spent 700,000 GPU hours finding universal jailbreaks to develop countermeasures, with a rapid-response process to fix new vulnerabilities.
Sponsored Protocol
US government request led OpenAI to limit preview to a small group of trusted partners
OpenAI provided the US government with a preview of GPT-5.6 before any public release. At the administration's request, the company restricted initial access to a small set of trusted partners whose participation was shared with the government. In its announcement, OpenAI stated: 'We don't believe this kind of government access process should become the long-term default.' The company views this as a temporary step to ensure a safe public rollout in the coming weeks. This move follows Anthropic's suspension of its Mythos and Fable models after a government directive due to jailbreak vulnerabilities. The US government has since allowed Anthropic to partially restore Mythos access for select organizations.
OpenAI's focus on jailbreak prevention likely stems from Anthropic's incident. Weeks ago, Anthropic blocked access to Mythos 5 and Fable 5 after Amazon and other companies reported that the models could be exploited. This episode accelerated the adoption of stricter security measures across the industry. For more on European implications, read the article on OpenAI Tightens Restrictions in Europe. Additionally, the vulnerability of model distillation was highlighted in a recent Anthropic report: Anthropic Reveals AI Model Distillation Vulnerability.
Sponsored Protocol
OpenAI plans to release GPT-5.6 fully in the coming weeks. For now, trusted partners are testing the three variants, while the industry watches the evolution of safety and cost in generative AI. According to a New York Times report, OpenAI, Anthropic, Google, xAI, and Microsoft had already granted the government early access to their latest models, while Meta remained a holdout, prompting the administration to urge the Menlo Park giant to submit its AI models for evaluation.
Source: https://www.engadget.com/2203102/openai-starts-previewing-gpt-56-and-its-three-variants
