OpenAI and the White House: GPT-5.6 Under Close Surveillance

OpenAI and the GPT-5.6 Series: A Controlled Advancement

OpenAI recently unveiled the GPT-5.6 series, succeeding GPT-5.5 with notable improvements in the fields of coding, biology, and cybersecurity. This new series consists of three distinct models: Sol, Terra, and Luna. However, at the request of the Trump administration, access to these models is currently restricted to a limited group of partners, the list of which has been communicated to the U.S. government. This approach could become a standard for future launches.

Historically, OpenAI named its models based on their usage, with designations such as Thinking, Instant, and Cyber. With the GPT-5.6 series, the company adopts a new system based on capabilities: Sol is the most powerful model, Terra is balanced, and Luna is designed to be fast and economical. These names are expected to remain consistent from one generation to the next.

The Sol model stands out for its unmatched performance at OpenAI. It introduces two new reasoning modes: the max mode, which allows the model to take more time to deepen its analyses, and the ultra mode, which utilizes multiple sub-agents to handle complex tasks simultaneously.

ChatGPT and Codex will soon integrate this series, following a preview phase reserved for selected partners. Sol will also be available via the Cerebras infrastructure, with speeds reaching 750 tokens per second, initially for a select group of clients.

Significant Advances in Multiple Areas

The GPT-5.6 Sol model presents notable improvements in three sectors compared to its predecessor GPT-5.5:

• Coding: Sol enhances command-line workflows, requiring planning, iterations, and tool coordination.

• Biology: The model is capable of processing complex genomic analyses with reduced token usage compared to its predecessor.

• Cybersecurity: Sol competes with Claude Mythos Preview from Anthropic while consuming about a third of the tokens in output.

Regarding cybersecurity, Sol does not meet the "Cyber Critical" threshold of OpenAI's security framework. In tests involving Chromium and Firefox, it was able to identify bugs and exploitation primitives but failed to produce a complete and functional exploit autonomously. OpenAI emphasizes that benchmark thresholds cannot cover all possible uses of a model or all combinations with other tools. This uncertainty justifies enhanced protective measures and the gradual rollout of models, limited to a few partners, a decision heavily influenced by the White House.

The Growing Influence of the U.S. Government

Before the launch, OpenAI presented its models to the Trump administration, which requested to limit the initial rollout. OpenAI collaborates with the executive branch while expressing its reservations: this process should not become the long-term norm, as it deprives various stakeholders of the best available tools. Nevertheless, OpenAI has not put up significant resistance to Washington's demands, which seems intent on generalizing this type of control.

OpenAI stated:

"We are taking this short-term measure because we believe it is the safest path toward broader availability in the coming weeks, while we work with the Administration to develop the framework for the presidential decree related to cybersecurity and a reproducible process for future model rollouts."

A few weeks ago, the government had already forced Anthropic to disable access to Claude Mythos 5 and Claude Fable 5 for its entire customer base, citing cybersecurity concerns and the potential access of China to the model. After two weeks of negotiations, the administration partially lifted the restriction and allowed Anthropic to redeploy Mythos 5 to over 100 U.S. institutions. However, Fable 5 remains disabled to this day.