OpenAI Limits GPT-5.6 Access After U.S. Request

For AI/ML engineers, security researchers, and enterprise teams, the core operational question is who gets early visibility into failure modes. Government-gated previews centralize adversarial testing inside a small number of approved partners - roughly 20 for GPT-5.6 and around 100 for Claude Mythos 5 - meaning the independent red-teaming, exploit-chain discovery, and robustness testing that normally surfaces edge cases before broad deployment is delayed or absent for the broader community. Benchmark results from this restricted window carry a structural bias: first-impression public figures come from vendor-reported or partner-selected evaluations rather than independent replications, giving vendors greater control over early performance narratives.

AP News and Axios report that OpenAI released three models in the GPT-5.6 family - Sol (flagship), Terra (balanced efficiency), and Luna (speed-optimized) - to a limited group of trusted partners on June 26, 2026, following a White House request citing cybersecurity concerns. This is the first time the U.S. government has preemptively asked an AI company to restrict a model release before launch, according to Axios and CNN. A parallel standoff involved Anthropic: the Commerce Department held its Fable 5 and Mythos 5 models pending review, then Commerce Secretary Howard Lutnick wrote to Anthropic clearing Claude Mythos 5 for release to approximately 100 companies and federal agencies, per CNBC and Semafor. AP reports the Trump administration cited concerns that the latest models could accelerate cyber vulnerabilities.

OpenAI reported GPT-5.6 Sol scored 88.8% on Terminal-Bench 2.1, narrowly ahead of Anthropic's Claude Mythos 5 at 88.0%; a new subagent "ultra" mode reached 91.9% (all vendor-reported figures). OpenAI separately stated that Sol's cybersecurity capabilities do not reach the "critical" threshold in its preparedness framework, framing restricted access as a precaution rather than a hard safety block. Independent verification of these benchmark figures is not yet possible given limited access; practitioners should treat the numbers as vendor claims until third-party replication is available.

OpenAI stated it "believes in broad access" and described the staggered release as "the strongest path to broader availability in the coming weeks," while also explicitly stating the government vetting process should not become the "long-term default" (CNBC, Forbes). The company said it is working with officials to develop a repeatable, documented framework for government model assessment. AP reports the Trump administration flagged accelerating cyber-vulnerability concerns as the driver of the request.

Whether the trusted-partner list is disclosed publicly; the timeline for broad GPT-5.6 availability (OpenAI cited "weeks"); whether independent benchmark results appear in public repositories and diverge from vendor-supplied numbers; how the government assessment framework takes formal shape; and whether Anthropic's Fable 5 receives similar clearance under the same process.