OpenAI has delayed the full public launch of GPT-5.6 after the US government requested early access and additional oversight before broader availability. The move signals a new era where frontier AI models face government审查 before reaching the public.
GPT-5.6 Public Rollout Paused
Initial access to GPT-5.6 is limited to a small group of vetted partners whose details were shared with authorities. The delay comes as Washington grows increasingly concerned that advanced AI systems could be misused for cyberattacks, military applications, or other national security threats.
OpenAI stated the delay is temporary while it works with the administration on a repeatable release process. However, the company warned that government control over customer access “should not become standard practice.” This sets a precedent that could affect how all major AI companies release future models.
Anthropic’s Models Caught in Export Controls
The US government also moved against Anthropic’s most advanced models. Claude Mythos 5 access was initially suspended over national security concerns, then partially restored to a limited group of over 100 trusted US organizations, many connected to critical infrastructure.
Anthropic described Mythos 5 as its strongest cybersecurity model, but access remains restricted for companies outside the approved list. The government has not clearly explained how organizations are selected or what safeguards changed to warrant the partial restoration.
Claude Fable 5: Restricted Then Reopened
In a parallel move, the Trump administration lifted export controls on Anthropic’s Claude Fable 5, restoring customer access after the model was pulled for security reasons 18 days earlier. The decision returned public access to a Mythos-class model, though the technical or policy changes Anthropic made to satisfy Commerce Department concerns remain unclear.
The administration faces an August deadline to create standardized benchmarks for assessing security risks in new AI models. This deadline could reshape how every AI company in the US plans and executes model launches.
New Safeguards Against Jailbreaks
Following the export control saga, the Commerce Department lifted controls on both Fable 5 and Mythos models after Anthropic implemented new safeguards designed to reduce jailbreak risks. Anthropic plans to expand access through its Glasswing program and is collaborating with Amazon, Google, and Microsoft on shared standards for identifying jailbreak techniques.
Anthropic acknowledged that completely eliminating jailbreaks may be impossible, but introduced protections that block risky behavior and route affected requests to more secure models. This collaborative approach between AI companies and the government could become the template for future model releases.
What This Means for AI Development
These government interventions create several new realities for the AI industry:
- Unpredictable release timelines: Marketing teams, product managers, and developers planning features around new AI capabilities must build in more flexibility around release timing.
- Uneven access: Frontier model access may become uneven across industries, vendors, and enterprise buyers, creating capability gaps that affect competitive positioning.
- Government review as standard: The August deadline for standardized security benchmarks could make government pre-release review a permanent fixture of AI development.
Major cloud providers and enterprise AI buyers should monitor whether these model restrictions, safety reviews, and shifting access rules change vendor capabilities and enterprise AI adoption plans.
Frequently Asked Questions
Why was GPT-5.6 delayed?
The US government requested early access to GPT-5.6 and additional oversight before OpenAI’s planned public rollout. The government is concerned that advanced AI systems could be misused for cyberattacks or military purposes.
Can anyone use Anthropic’s Mythos 5 model?
No. Mythos 5 access is restricted to over 100 trusted US organizations, many connected to critical infrastructure. Companies outside the approved list cannot access the model, and the government has not disclosed the selection criteria.
What is Anthropic’s Glasswing program?
Glasswing is Anthropic’s program for expanding access to its AI models. The company is using it alongside collaborations with Amazon, Google, and Microsoft to establish shared standards for identifying and addressing jailbreak techniques across the industry.
When will the government set AI security benchmarks?
The Trump administration faces an August 2026 deadline to create standardized benchmarks for assessing security risks in new AI models. These benchmarks could make government pre-release review a standard part of AI model launches going forward.
