Frontier-model access is becoming staged and approval-sensitive. That does not mean the United States has created a universal licensing regime for AI models.

The mechanisms differ. OpenAI says GPT-5.6 is in a small trusted-partner preview whose participation was shared with the government. Anthropic says the separate export controls on Fable and Mythos were lifted June 30: Fable returned globally July 1, while Mythos returned only to a set of approved US organizations.

Current Access Matrix

Status checked July 1, 2026. Console access remains the operational source of truth for a specific account, plan, cloud, and region.

ModelAnnouncedPreview-accessibleAccount-approvedGenerally availableCurrent read
GPT-5.6 SolYesYes, API and CodexProvider-selected cohort; verify your organizationNoOpenAI-confirmed limited preview; broader availability is planned, not delivered
GPT-5.6 TerraYesYes, API and CodexProvider-selected cohort; verify your organizationNoSame preview gate as Sol
GPT-5.6 LunaYesYes, API and CodexProvider-selected cohort; verify your organizationNoSame preview gate as Sol
Claude Fable 5YesYesVerify plan, credits, cloud, and regionYes on native Claude surfacesAnthropic restored global access July 1; included usage is temporary and cloud re-enablement is rolling out
Claude Mythos 5YesYes, approved US organizationsTrusted-access eligibility requiredNoAnthropic restored access to a set of approved US organizations and is coordinating broader Glasswing access

Source boundary: OpenAI’s own wording is “trusted partners whose participation has been shared with the government.” The claim that the launch cohort was approximately 20 companies whose participation was approved by government is Axios reporting, not OpenAI’s published wording. Fable and Mythos status now comes from Anthropic’s June 30 replacement update; the earlier AP and Business Insider reports remain timeline evidence.

Dated Timeline

DateSource-labeled event
June 2, 2026The White House executive order directed agencies to design a voluntary framework for review of covered frontier models, including up to 30 days of government access before release to other trusted partners. It explicitly says this does not authorize mandatory governmental licensing, preclearance, or permitting.
June 9, 2026Anthropic announced Fable 5 as broadly available and Mythos 5 as a Project Glasswing/trusted-access model. Its live platform docs still preserve that launch-state language.
June 12, 2026Anthropic said it was removing Fable 5 and Mythos 5 access for all users to comply with a US government directive. This was a vendor statement about a specific directive, not proof of a standing licensing rule for every frontier release.
June 24, 2026Business Insider reported that Fable had returned behind nationality-based controls and enhanced onboarding. Anthropic’s suspension page remained live, so AIHackers treated the report as partial, account-specific evidence rather than general availability.
June 26, 2026OpenAI announced a trusted-partner preview for GPT-5.6 Sol, Terra, and Luna and said broader availability was planned. OpenAI also said it did not want this government access process to become the long-term default.
June 26, 2026AP reported that Mythos 5 could be redeployed to a small group of cyber defenders and infrastructure providers. AP separately reported that Fable was still unavailable and that Anthropic intended to keep working toward broader access.
June 30-July 1, 2026Anthropic said the export controls were lifted, restored Fable globally on native Claude surfaces from July 1, and restored Mythos to a set of approved US organizations. Cloud-provider re-enablement remains in progress.

Different Gates, Not One Regime

ControlGPT-5.6 Sol / Terra / LunaFable 5 / Mythos 5
Government reviewOpenAI says it previewed plans and capabilities to government, then accepted a temporary limited rollout requestAnthropic says a separate directive forced an all-user suspension; the public directive text is not available here
Trusted-access programInitial OpenAI preview cohort; participation shared with governmentFable is broadly restored; Mythos remains a Glasswing/approved-organization path
Account eligibilitySelect trusted partners and organizations; OpenAI has not published a self-serve eligibility contractFable depends on plan, credits, cloud, and region; Mythos requires trusted-access approval
Nationality controlsOpenAI’s announcement does not publish a nationality rule for the previewAnthropic says the June restrictions were lifted; Mythos still has organization-level approval
Data retentionThe GPT-5.6 announcement does not publish a model-specific retention exception; verify the applicable account contractAnthropic docs specify 30-day retention and no zero-data-retention option for Fable/Mythos
SafeguardsOpenAI documents layered cyber safeguards, possible blocks, delays, and review during previewFable can refuse or fall back in guarded domains; Mythos removes some Fable classifiers for approved cohorts
Fallback behaviorOpenAI warns legitimate work can be blocked or delayed; the launch post does not document an automatic fallback modelAnthropic documents HTTP 200 refusals and server-, client-, or manually implemented fallback to another Claude model

The June 2 order matters, but its text is narrower than “government licensing.” It calls for a voluntary framework, classified benchmarking, protected government access for up to 30 days, and collaboration on trusted partners. It also expressly rejects reading that framework as mandatory licensing or preclearance.

The June Anthropic incident is more coercive in the company’s telling: Anthropic says it received a legal directive and removed access for all users. The public record does not establish that this mechanism automatically applies to every company or every model.

What Teams Should Verify Before Production

  1. Console access: confirm the exact model ID is visible and callable in the intended API, Codex, cloud, or first-party console.
  2. Compliance eligibility: record the organization, account, region, nationality, trusted-program, and end-user constraints that apply.
  3. Data retention: obtain the model-specific retention and zero-data-retention terms in writing. Do not infer them from another model family.
  4. Guardrail behavior: run representative cyber, biology, code-review, and dual-use prompts; log refusals, delays, classifier signals, and false positives.
  5. Fallback models: verify whether fallback is automatic, which model receives the prompt, whether context is retained, and how retries are billed.
  6. Production alternatives: keep GPT-5.5 and Claude Opus 4.8 as active premium baselines, plus an approved lower-cost or independent-provider route.

Archive Evidence and Source Conflicts

Policy-sensitive primary pages were checked through the Wayback CDX workflow. Each link below is a direct 200 text/html capture; timestamps are UTC.

Status transition preserved: the launch, suspension, press reports, and June 30 redeployment page describe different points in the incident timeline. The latest primary source supports Fable general availability on native Claude surfaces from July 1, while Mythos remains approved-organization access and cloud-provider restoration is incomplete.

Sources

Last verified: July 1, 2026. Fable is restored globally on native Claude surfaces; Mythos and GPT-5.6 remain approval-sensitive. Plan allowances, cloud rollout, account approvals, and vendor documentation can change independently.