OpenAI Restricts Access to GPT-5.6 at the Request of the US Government

Access via the API and Codex is initially granted only to selected partners. "This step is being taken in close coordination with the US government," OpenAI said. The company announced that it would be made available "to the broader public via ChatGPT, Codex, and the API in the coming weeks." "Currently, seeing the agreement with this as the best way to expand access in the coming weeks," OpenAI said.

Three Model Variants and New Features

According to OpenAI, the API prices for the Terra variant are $2.50 per one million input tokens and $15 per one million output tokens. Access to the high-speed version of the Sol model on Cerebras hardware, which is expected to enable speeds of up to 750 tokens per second starting in July, is "likewise initially limited to selected customers." "According to the manufacturer, the Terra variant should offer comparable performance to the previous model GPT-5.5, while costing only half as much," OpenAI explained.

GPT-5.6 introduces two new features. "Max Reasoning Effort: A function that gives the model more time to thoroughly analyze complex problems," OpenAI explained. In addition, there is "Ultra Mode: A mode that accelerates the handling of complex tasks by deploying sub-agents that go beyond the capacity of a single agent." In tests such as Terminal-Bench 2.1 for command-line workflows and GeneBench v1 for genomics analyses, Sol achieved new best scores.

To protect the models, OpenAI "implemented a multi-layered security system. This includes safeguards directly in the model training, real-time checks during generation, as well as account-level controls." For the review of potential vulnerabilities, "over 700,000 A100 GPU hours for automated red-teaming" were spent. "The company emphasizes that the models were trained to refuse unauthorized assistance with cyberattacks, while legitimate tasks such as security research or bug fixing (patch development) should continue to be supported."

Security System and Missing External Tests

Independent evaluations are still pending: "Due to the restrictions, there are no independent evaluations and benchmarks, neither at Arena.ai nor at Artificial Analysis." The goal is "to review the models' capabilities before a wider release and to work together with the authorities on a framework for the security of cybersecurity models." OpenAI also stated that it does not consider such government access restrictions to be a desirable permanent solution.

In parallel with the release of GPT-5.6, competitor Anthropic is generating discussion. Its new most powerful model was withdrawn again a few days after release at the request of the US government, because it is particularly good at finding software vulnerabilities. In the publicly available version, the corresponding capabilities were blocked.

Austria Lobbies for Anthropic in Europe

In Austria, the State Secretary for Digitalization, Alexander Pröll (ÖVP), responded with a letter to EU Commissioner Henna Virkkunen. He proposed a strategic establishment and participation of Anthropic in the EU. "Let us jointly examine a strategic establishment and participation of Anthropic within the European Union. With legal certainty, market access, capital, and a value system that fits this company," the letter states.

Pröll justified the initiative to the APA by saying he wanted to strengthen the location and become more digitally sovereign. "Especially Anthropic is a fit for us. A company that does not view the ethical use of AI as marketing, but as a fundamental conviction. That puts safety above speed. That is a deeply European attitude. This company would not be constrained in Europe; it would be set free," he said. "One must not close one's eyes to this reality," Pröll added with regard to the US restrictions.

At the same time, Pröll warned of supply risks: "Europe must not be cut off from cutting-edge innovations." He made clear, however: "It is not about undermining our own European companies and models."