GPT-5.6: A Frontier Model You Can’t Freely Use
GPT-5.6 is a new generation of frontier AI models from OpenAI that promise major advances in coding, biology, and security tasks, yet their deployment is restricted to a small, government-approved group of users while wider developers and everyday customers wait for access.
OpenAI has introduced a family of GPT-5.6 models—Sol, Terra, and Luna—designed to span high-end reasoning, balanced work, and low-cost, high-volume tasks. Sol is described as “competitive” with Anthropic’s Mythos Preview for finding vulnerabilities while using roughly a third of the output tokens, a clear signal that OpenAI has pushed efficiency and power at the same time. Terra aims to deliver performance comparable to GPT-5.5 at half the cost, and Luna targets the lowest-cost scenarios without abandoning modern capabilities. On paper, this is the logical next step for AI adoption: better security, better coding scores than Mythos 5, and more efficient biology workflows. In practice, most users only see a locked door. The top tier of AI innovation has shifted from a consumer launch story to a controlled, state-influenced preview.
Government-Driven GPT-5.6 Access Restrictions
The most significant fact about GPT-5.6 is not its capability; it is its gatekeeper. OpenAI confirms that preview access is limited “at the request of the US government,” and that it first demonstrated Sol, Terra, and Luna to federal officials before launch. As part of a previous agreement that lets the Defense Department use OpenAI’s models, the company is starting with a “small group of trusted partners” whose identities are shared with the government before any wider rollout.
Behind the scenes, the Office of the National Cyber Director and the Office of Science and Technology Policy reportedly worked closely with OpenAI and pushed for a restricted initial release. According to one report, federal officials will approve GPT-5.6 use “customer by customer” in the early phase—a stark example of OpenAI government limitations shaping who sees frontier AI first. This is frontier AI regulation by access throttle: no explicit licensing regime, but real control through GPT-5.6 access restrictions.
A Pattern: Frontier AI Regulation by Backdoor
GPT-5.6 does not exist in a vacuum; it follows a growing pattern of government restrictions on frontier AI models. Anthropic’s Mythos Preview was initially limited to a small circle of cybersecurity firms and trusted companies due to its advanced hacking capabilities, and that disclosure set off a surge in AI model safety debates among investors and political leaders. When Anthropic later released Fable 5, it pitched the model as “Mythos-class” but with stronger safeguards, yet even that did not prevent heavy scrutiny.
Then came export controls that forced Anthropic to revoke access to Mythos and Fable entirely on June 12, after new federal rules took effect. In parallel, a June 2 executive order called on AI developers to voluntarily submit advanced models for government review for up to 30 days before public release—a mechanism explicitly aimed at security, intellectual property, and confidentiality risks. The GPT-5.6 rollout mirrors Anthropic’s Project Glasswing, where early Mythos access was limited out of fear it was too powerful. The lesson is clear: frontier AI regulation is emerging not only through formal law, but through quiet, model-by-model bargaining over who gets early access.
What Users Lose When ChatGPT Deployment Delays Become the Norm
For ordinary users and enterprises, GPT-5.6 is technically ready but functionally absent. At the moment, OpenAI is offering the models only through its API and Codex to a “select group” of trusted organizations, with broader availability via ChatGPT, the API, and Codex promised “soon.” Officially, this limited preview helps test whether safeguards can block harmful use without choking legitimate work; feedback is supposed to refine context understanding and reduce arbitrary refusals.
OpenAI has signaled discomfort with this arrangement, warning that such government access processes “keep the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” Meanwhile, the pricing signals how central GPT-5.6 could become: Sol is set at USD 5 per input and USD 30 per output (approx. RM23 and RM138), Terra at USD 2.50 per input and USD 15 per output (approx. RM11 and RM69), and Luna at USD 1 per input and USD 5 per output (approx. RM4.60 and RM23). Those ChatGPT deployment delays do not only slow experimentation; they postpone concrete productivity gains in coding, genomics, and security that customers are willing to pay for.
Security First, But at What Strategic Cost?
Supporters of tight GPT-5.6 access restrictions argue that models capable of competing with Mythos Preview on vulnerability discovery must be treated as dual-use technology. Sol includes layered safeguards, jailbreak checks, and the ability to block suspicious prompts even when they look benign. Seen through that lens, staged deployment is caution, not control. The voluntary review window created by the June 2 executive order encourages companies to seek what amounts to a soft endorsement in exchange for cooperation.
The risk is that a temporary safety check becomes a permanent bottleneck. OpenAI has already hinted that it does not want this government access process to be the default for every frontier model release. If each new generation must pass a vague, customer-by-customer approval process, AI adoption will skew toward well-connected institutions and away from smaller firms, public defenders, and open research labs. The long-term danger is not that frontier AI regulation exists, but that it quietly hardcodes inequality into who can experiment with, and benefit from, the most advanced tools.
