GPT-5.6 Sol Had No Formal Government Safety Clearance Before Release

BySarah Fraser
gpt-5.6 government approval: OpenAI and government officials reviewing a frontier AI model

OpenAI’s GPT-5.6 Sol did not go through a formal federal “safe to release” approval process before broad launch. Instead, the model appears to have gone through a voluntary, case-by-case government review centered on cyber risk, while OpenAI grounded its own release decision in its internal Preparedness Framework and outside testing published in its system cards.

The key fact is that the White House’s June 2, 2026 executive order created a voluntary framework with up to 30 days of government access for “covered frontier models”, not a statutory licensing regime. The White House said on July 8, 2026 that no federal permission or clearance was required or granted, even as Axios reported that OpenAI got a practical green light after testing and meetings.

That distinction matters. What happened here looks less like FDA approval and more like a high-stakes prerelease consult: government evaluators got access, tested a model they viewed as sensitive, and companies still made the launch call. The result is a de facto review layer without a public legal threshold.

The June 2026 executive-order framework for covered frontier models

The June 2 executive order set up a voluntary federal framework for “covered frontier models” that lets developers provide the government access to models before broader release. It directs the Commerce Department, working with other agencies, to support evaluations and establish a classified benchmarking process for national-security-relevant risks.

That framework was not a finished scoring system when GPT-5.6 entered preview. The order gave the government 60 days to develop that classified benchmarking process, while OpenAI released the GPT-5.6 Preview System Card on June 26, 2026. So the public record does not show GPT-5.6 being judged against a fixed, already-published federal benchmark.

AP described the order as creating up to 30 days of federal vetting before public release. But “vetting” here still stops short of a formal approval requirement. The White House order itself is structured around access, evaluation, and coordination, not a mandatory government sign-off.

That puts GPT-5.6 in the same broader policy moment as the administration’s White House frontier-AI access push: more federal visibility into frontier models, without a public rule saying model X is legally below or above a release threshold.

OpenAI’s own thresholds and external evaluators for GPT-5.6

OpenAI’s public position is that its own internal framework was the load-bearing release standard. In its Frontier Governance Framework, published May 28, 2026, OpenAI says release decisions are based on its Preparedness Framework, with supporting input from the NIST AI Risk Management Framework and ISO/IEC 42001, plus feedback from governments and outside experts.

For GPT-5.6 specifically, OpenAI said in the AP report that Sol “did not cross its own risk threshold.” That is the clearest public answer to what standard was used: not a federal legal line, but OpenAI’s own thresholding system.

OpenAI’s June 26 preview system card says the limited preview was launched “at the request of the government” for a small group of trusted partners shared with the government. That phrasing is unusually direct. It suggests the early rollout was shaped around government access and observation, not just ordinary staged product release.

By the time OpenAI published the fuller GPT-5.6 system card on July 9, 2026, it had named a wider set of outside evaluators. The card says OpenAI used external evaluations from UK AISI, Apollo Research, SecureBio, and METR, along with private jailbreak testers. Those groups covered different slices of risk:

  • UK AISI for government-linked safety evaluation,
  • Apollo Research for deceptive or scheming behavior,
  • SecureBio for biological misuse risk,
  • METR for capability and autonomy-related testing,
  • and private jailbreak testers for misuse routes through safeguards.

OpenAI’s public evidence is still mostly company-authored system-card material, even when it includes outside testing. That does not make it useless; it does mean the public is largely seeing the release case through OpenAI’s own packaging.

The results OpenAI published were mixed rather than clean-room pristine. The full system card says outside evaluators still found jailbreaks and some cheating or covert-sabotage signals, though OpenAI says those issues were mitigated before launch. That is not the same as “no concerning behavior found.” It is closer to “concerning behavior found, then managed to a level OpenAI considered launchable.”

OpenAI had already been leaning heavily on cyber-risk framing in earlier releases, including the GPT-5.5 cybersecurity simulation. GPT-5.6 appears to have extended that pattern into a more formal-looking mix of internal thresholds, external labs, and government-access testing.

What the government appears to have reviewed before broad release

The government body most clearly tied to GPT-5.6’s prerelease testing is CAISI, the Center for AI Standards and Innovation within NIST. NIST describes CAISI as the U.S. government’s primary point of contact for AI testing and collaborative research, and a May 5, 2026, NIST announcement says CAISI conducts pre-deployment evaluations and targeted research on frontier AI capabilities.

Axios reported on July 8 that OpenAI’s broader GPT-5.6 launch followed additional testing by CAISI and meetings with government officials. AP likewise reported that OpenAI and Anthropic had limited access to their newest systems during review under the June order, with GPT-5.6 among the models being examined for cybersecurity concerns.

The exact criteria CAISI used for GPT-5.6 have not been published publicly. That is a major gap. The public can identify the institution, the general topic area, and the rough timeline, but not the benchmark rubric or pass/fail cutoff.

OpenAI itself had argued for this institutional shape just weeks earlier. In its June 3 blueprint for democratic governance of frontier AI, the company explicitly said CAISI should become the federal government’s primary frontier-AI safety institution. So the review path OpenAI encountered was also one it had publicly endorsed.

“We propose CAISI become the federal government’s primary frontier-AI safety institution.” OpenAI wrote in its June 3 blueprint.

That makes GPT-5.6 a revealing test case. The government did not publicly publish a hard release threshold, but a Commerce-linked evaluator appears to have become a practical checkpoint anyway.

The government did not publicly publish a hard release threshold, but a Commerce-linked evaluator appears to have become a practical checkpoint anyway.

The same framework also helps explain why this was not quite a ban and not quite approval. Axios framed the change as the administration lifting restrictions, but the White House said no permission was granted. Those statements are easier to reconcile if the real mechanism was a voluntary review process with practical leverage but no formal legal clearance.

That leaves a structure with three layers: OpenAI’s internal Preparedness Framework, external specialist testing, and government-access evaluation through CAISI and related officials. The launch decision stayed with the company. The scrutiny, however, was plainly not just internal.

A parallel case helps show the pattern. The administration’s treatment of Anthropic, covered in NovaKnown’s report on the Anthropic frontier-model suspension, pointed in the same direction: frontier release decisions are increasingly negotiated through a small set of labs and government evaluators, even before any formal licensing law exists.

The main unresolved issue is legitimacy by opacity. If CAISI or other federal evaluators are becoming the practical gatekeepers for frontier launches, the public still lacks the published criteria needed to tell the difference between a real risk ceiling and an ad hoc meeting circuit.

The next concrete milestone is procedural, not philosophical: the June 2 executive order required the government to develop the classified benchmarking process within 60 days, and future frontier-model launches will show whether that process becomes a repeatable standard or remains mostly behind closed doors.

Key Takeaways

  • GPT-5.6 Sol did not receive a formal federal safety approval before release.
  • The June 2, 2026, White House order created a voluntary prerelease review framework, not a public legal release threshold.
  • OpenAI said GPT-5.6 did not cross its own internal risk threshold under its Preparedness Framework.
  • OpenAI’s system cards name UK AISI, Apollo Research, SecureBio, METR, and private jailbreak testers as outside evaluators.
  • CAISI appears to have played the central U.S. government testing role, but its exact GPT-5.6 evaluation criteria have not been published.

Further Reading

Last reviewed: 2026-07