OpenAI has begun showing off its GPT-5.6 series to a small circle of trusted partners, with a wider release planned for the coming weeks. The new series comes in three versions: Sol, Terra, and Luna, each aimed at a different type of user and use case. As reported by Engadget, the company positioned Sol as its most capable model to date, while Terra and Luna target cost-conscious users who do not need maximum performance.
The rollout is not entirely business as usual. OpenAI gave the US government a preview of GPT-5.6 before making it available to partners, and the partner preview itself is happening partly because the administration requested it. The company says partners in the preview are ones “whose participation has been shared” with the government. OpenAI was quick to note this is a short-term arrangement. “We don’t believe this kind of government access process should become the long-term default,” the company wrote in its announcement.
That context matters. President Trump signed an AI cybersecurity order earlier this month asking companies to submit their most powerful models for voluntary government review 30 days before any public release. According to a recent New York Times report, OpenAI, Anthropic, Google, xAI, and Microsoft had already been giving the government early access to their models even before the order was signed. Meta was the only holdout, and the US government has reportedly been pushing it to follow suit.
The three variants in the GPT-5.6 series are clearly differentiated by capability and price:
- Sol is the flagship model, built for deep reasoning and cybersecurity work. It includes a new “max” reasoning effort mode that gives it more time to think through complex problems. OpenAI says Sol is its best model for finding and fixing software vulnerabilities.
- Terra is designed for everyday tasks. OpenAI says it performs similarly to GPT-5.5 but costs half as much.
- Luna is the cheapest option in the series, aimed at high-volume, lower-stakes tasks.
Pricing for the series reflects that tiered structure. Sol costs $5 per million input tokens and $30 per million output tokens. Terra comes in at $2.50 for input and $15 for output. Luna is the most affordable at $1 for input and $6 for output. For comparison, Anthropic’s Fable model, before it was pulled from availability, cost $10 per million input tokens and $50 per million output tokens, making Sol a notably cheaper option for high-end AI work.
Security is a major theme running through the GPT-5.6 launch. OpenAI says it spent several weeks stress-testing the models before release, looking for weaknesses and hardening them against real-world attacks. The company also trained all three variants to refuse what it calls “prohibited cyber assistance,” which includes attempts to jailbreak the model. It dedicated 700,000 GPU hours to finding universal jailbreaks and developing defenses against them. OpenAI also says it will run a rapid-response process to identify and address newly discovered jailbreaks going forward.
That focus on jailbreak prevention comes at a sensitive time for the industry. A few weeks ago, Anthropic suspended access to its Mythos 5 and Fable 5 models following a government directive. The company did not officially explain why, but reports indicated that Amazon and other companies had alerted authorities that the models could be jailbroken for malicious use. Access has since begun to be restored after the US government gave Anthropic permission to redeploy Mythos to a select group of organizations. OpenAI’s aggressive approach to jailbreak prevention looks like a direct response to that situation, an attempt to avoid a similar outcome with its own models.
The broader picture here is that AI companies are operating in a new regulatory environment where government oversight is no longer theoretical. The preview process for GPT-5.6 shows how quickly that relationship between the AI industry and Washington has shifted, and how much companies like OpenAI are now building their release strategies around it.
