The US government is now a customer gatekeeper for both OpenAI and Anthropic's frontier models: GPT-5.6 Sol and Claude Mythos 5 ship under the same Commerce Department vetting regime

On 2026-06-26, the same day, two related things happened at the US frontier-AI frontier. OpenAI began a limited preview of GPT-5.6 Sol with the customer list “shared with the government” (OpenAI, 2026-06-26). On the same day, the Commerce Department lifted its two-week-old export block on Anthropic’s Claude Mythos 5, allowing it to be released to “more than 100 US institutions, including major companies and government agencies” via a letter from Commerce Secretary Howard Lutnick to Anthropic’s chief compute officer Tom Brown (Semafor, 2026-06-26, 6:30pm EDT). Together the two actions establish a new distribution model in which each new customer of a US frontier lab’s most capable model is, in practice, approved by the Trump administration’s Commerce Department. Semafor frames the Lutnick letter as “the beginnings of a new regulatory regime that gives the US government control over the release of frontier AI models,” and “the framework for overseeing AI is being built on the fly.” The Washington Post frames the move as “expanding its recent policy of vetting companies that want access to the latest artificial intelligence technology.” No Federal Register entry, no executive order, no published criteria. OpenAI, in its own announcement, is publicly pushing back: “We don’t believe this kind of government access process should become the long-term default.”

The policy is the story. The model releases are evidence.

What happened — the policy regime, evidenced by two releases

OpenAI’s announcement describes the GPT-5.6 Sol preview as “a limited preview for a small group of trusted partners whose participation has been shared with the government” (OpenAI, 2026-06-26). The Lutnick letter, as quoted by Semafor, grants a license exemption for “entities identified in Annex A to this letter and their foreign national employees, or to Anthropic’s foreign national employees” (Semafor, 2026-06-26). The Washington Post account of the letter says the government reserves the right to “change the list of companies at any time.” The Annex A customer list is not public.

The structural shift is the new customer-vetting regime, applied to both labs. Per WaPo, “The Trump administration is requiring both Anthropic and OpenAI to get approval for each new customer of their most powerful AI technology” (WaPo, 2026-06-26). The word “both” is doing work: this is no longer a one-off Mythos export-control case but a policy of pre-approval for frontier customers, applied across the two US frontier labs. WaPo’s framing makes the expansion explicit — “The Trump administration is expanding its recent policy” — and notes the OpenAI intervention “was the first time it had extended government vetting of AI customers beyond Anthropic.”

The regime is being built without public rules. Semafor: “The framework for overseeing AI is being built on the fly, and many users of the powerful tools — from non-US governments and companies to consumers — remain in the dark as to when they will get access to Mythos and Fable” (Semafor, 2026-06-26). No published criteria, no public rule, no Federal Register entry, no executive order cited in any of the three load-bearing sources. OpenAI names a “cyber Executive Order framework” as the path out of the current ad-hoc arrangement, but no such order has been published as of 2026-06-27 (OpenAI, 2026-06-26). The Commerce Department characterises the process as fast and on track — spokesman Benno Kass: “In just two weeks, we have worked diligently to ensure America remains the global leader in AI while safeguarding our security” (Semafor, 2026-06-26).

OpenAI is publicly pushing back, in its own announcement. “We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks, while we work with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases” (OpenAI, 2026-06-26). CEO Sam Altman is more pointed on X: “I just dont like the idea of the government picking the customers … Confident we will get to a better place” (WaPo, 2026-06-26).

Allied reaction: frustration at US dependency. Per Semafor, “European officials and other US allies have expressed frustration at their new dependence on decisions in Washington” (Semafor, 2026-06-26). WaPo adds that British MP Kanishka Narayan said Britain’s AI Security Institute had access to OpenAI’s GPT-5.6 — “it is the only non-U.S. entity with access to it” per a person familiar with the matter (WaPo, 2026-06-26).

The two-week backstory. Per Semafor, “Two weeks ago the administration imposed export controls on Mythos, leading to a shut down of the model and its cousin Fable 5 after warnings from Amazon and other companies that they could be ‘jailbroken’ for malicious purposes” (Semafor, 2026-06-26). The 2026-06-13 Semafor exclusive on the SK Telecom Chinese-access concern is the named trigger; the Wired piece is the supporting link. AI Newsroom covered the 2026-06-13 suspension (AIN-60) and the policy framework (AIN-61) in earlier articles; this article builds on those, not in place of them.

What happened — the model-release evidence

GPT-5.6 series launches with a new naming scheme. Per OpenAI: “We’re beginning a limited preview of the GPT-5.6 series: Sol, our flagship model; Terra, a balanced model for everyday work; and Luna, a fast and affordable model. Terra has competitive performance to GPT-5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost” (OpenAI, 2026-06-26). The naming system: “the number identifies a model’s generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence.”

Pricing per 1M tokens and the cache change. Per OpenAI: “Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output” (OpenAI, 2026-06-26). On caching: “GPT-5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. For GPT-5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount.”

New reasoning effort and a new “ultra” mode. Per OpenAI: “With GPT-5.6, we’re introducing a new max reasoning effort to give Sol the most time to reason deeply. Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work” (OpenAI, 2026-06-26).

Capability claims — all vendor-attributed. OpenAI’s announcement makes four capability claims, each OpenAI’s own. Terminal-Bench 2.1: “GPT-5.6 Sol sets a new state of the art … which tests command-line workflows requiring planning, iteration, and tool coordination” (OpenAI, 2026-06-26). GeneBench v1: “it achieves stronger results than GPT-5.5 while using fewer tokens.” ExploitBench²: “GPT-5.6 Sol is competitive with Mythos Preview using only ~1/3 of the output tokens” — a cross-vendor claim the article attributes to OpenAI, not asserts. ExploitGym, “created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs”: “GPT-5.6 Sol, Terra, and Luna models all demonstrate strong improvements in cyber capabilities as we increase reasoning.” OpenAI also reports “over 700,000 A100-equivalent GPU hours … to automated red teaming aimed at finding universal jailbreaks.”

Cyber Critical threshold and the safeguard stack. Per OpenAI: “GPT-5.6 Sol does not cross the Cyber Critical threshold under our Preparedness Framework. In evaluations involving Chromium and Firefox, it identified bugs and exploitation primitives — the building blocks of an exploit — but did not autonomously produce a functional full-chain exploit under the conditions tested” (OpenAI, 2026-06-26). OpenAI’s own threshold, OpenAI’s own assessment, OpenAI’s own caveat — “benchmark thresholds cannot capture every way a model may be used or combined with other tools.” The layered safeguard stack is “protections trained into the model, real-time checks during generation, account-level signals, differentiated access, monitoring, enforcement, and continued testing.”

The Mythos 5 letter, in Semafor’s quoted text. Commerce Secretary Howard Lutnick, in a Friday letter to Anthropic’s chief compute officer Tom Brown: “I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model” (Semafor, 2026-06-26). Lutnick also wrote: “Anthropic has committed to work with the U.S. government on protocols and standards and releases.” The Fable 5 question is unresolved: “The letter is silent on Fable 5 … People close to the talks said they are moving toward releasing Fable as well, though that timeline is unclear.” Fable was previously a public model; Mythos was not.

Why it matters now

It is a structural shift, not an event. Two frontier labs are now under the same Commerce Department customer-vetting regime, with the same day as the visible evidence. The story is the regime, not the model releases. WaPo’s word for it is “expanding” — the policy is the expansion of a pre-existing process, and the two same-day announcements are the visible evidence that the expansion is now in effect at the two US frontier labs (WaPo, 2026-06-26).

The regime is being built without public rules. No Federal Register entry, no executive order, no published criteria. “Built on the fly” is Semafor’s direct framing, and it is the load-bearing critique (Semafor, 2026-06-26). Rep. Sam Liccardo (D-California) put it directly to WaPo: “The most kleptocratic regime in U.S. history has positioned itself as the unreviewable gatekeeper of the world’s most powerful technology, and nobody inside or outside the administration can say what standards it’s using to decide which trillion-dollar companies it’s allowing through the gate” (WaPo, 2026-06-26). Former Trump AI adviser Dean Ball, joining OpenAI next month, wrote: “In a matter of weeks, U.S. federal AI policy has gone from implausibly libertarian to increasingly draconian and opaque.”

OpenAI is publicly unhappy, in its own announcement. The most useful single line in the entire story for a builder audience. OpenAI is the lab asking the loudest for the regime to end, in its own announcement (OpenAI, 2026-06-26).

The customer lists are not public. WaPo notes a White House official said the administration “signed off on a list of companies OpenAI asked to be allowed access to Sol but excluded a handful of entities located outside of the United States” and that there is “no process for individual users to get access to the new model” (WaPo, 2026-06-26). Semafor notes the Annex A list is not public (Semafor, 2026-06-26).

Foreign users, including US allies, are now dependent on US government decisions. Semafor’s “European officials and other US allies have expressed frustration at their new dependence on decisions in Washington” is the named second-order effect (Semafor, 2026-06-26). The British AI Security Institute as the only non-U.S. entity with access to GPT-5.6, and the G7 concerns from France and Canada, are the documented signals (WaPo, 2026-06-26).

The cyber Executive Order is the named next step. OpenAI’s “we work with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases” makes the Executive Order a known unknown (OpenAI, 2026-06-26). The article treats the Executive Order as the main near-term regulatory watch item and does not predict its content.

Practical implications for builders and operators

Frontier-model procurement is now a regulatory event, not a sales event. Until the cyber Executive Order lands, the US government is the de facto customer-vetting step for OpenAI Sol and Anthropic Mythos 5. Per WaPo: there is “no process for individual users to get access to the new model” (WaPo, 2026-06-26).

GPT-5.6 Terra and Luna are NOT under the customer-vetting regime, per OpenAI’s announcement. The customer-vetting is for the “limited preview” of Sol “for a small group of trusted partners whose participation has been shared with the government” (OpenAI, 2026-06-26). Terra and Luna are described as additional models in the same series, but the announcement does not put them under the same step.

Cache economics change for GPT-5.6. “Cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount” (OpenAI, 2026-06-26). The 30-minute minimum cache life and explicit cache breakpoints are new levers worth modelling before any production migration.

Do not plan production workloads on the GPT-5.6 Sol preview. The preview is by design limited, the customer list is government-coordinated, and OpenAI explicitly says broader availability is “in the coming weeks” — not a calendar date (OpenAI, 2026-06-26). OpenAI also notes a Sol variant on Cerebras “at up to 750 tokens per second in July” with limited access — a separate track.

For non-US and non-Annex-A companies, the Mythos 5 question is the regulatory event of the week. Per WaPo, the Commerce Department’s letter to Anthropic said “only U.S.-based companies have access” (WaPo, 2026-06-26).

Risks and caveats

highRiskClaims: true is set. The picture that makes this story worth covering is what makes the caveats load-bearing.

1. All GPT-5.6 capability claims are OpenAI’s. Terminal-Bench 2.1 SOTA, GeneBench v1 stronger-than-5.5, ExploitBench² “competitive with Mythos Preview” at ~1/3 output tokens, ExploitGym improvements, the 700,000 A100-equivalent GPU hours — every one is OpenAI’s own measurement. The article attributes; it does not assert.
2. The “competitive with Mythos Preview” claim is a cross-vendor claim attributed to OpenAI. The article does not present it as an independent result.
3. “Does not cross the Cyber Critical threshold” is OpenAI’s self-assessment under OpenAI’s own Preparedness Framework. No independent body has confirmed the assessment.
4. The “new regulatory regime” is Semafor’s framing, confirmed by WaPo and the OpenAI announcement. The article may state that “the US government is now a customer gatekeeper for both OpenAI and Anthropic” because three primary sources support it, but does not assert a more specific legal characterisation.
5. No published rule, no executive order, no Federal Register entry. The article does not infer a process, a rule, or criteria that is not in one of the three primary sources. The cyber Executive Order is named as a path forward by OpenAI; the article does not characterise the order beyond OpenAI’s own framing.
6. The Annex A customer list for Mythos 5 is not public. The article does not name specific customers. “More than 100 US institutions, including major companies and government agencies” (Semafor) is the strongest form permitted by the sources.
7. The Fable 5 status is unresolved. Semafor: “moving toward releasing Fable as well, though that timeline is unclear.” The article does not state a date.
8. “European officials and other US allies have expressed frustration” is Semafor’s paraphrase, not a quote. The article attributes, and does not name specific officials or specific statements.

What to watch

1. The cyber Executive Order itself. OpenAI’s own announcement names it as the path out of the current ad-hoc regime.
2. Fable 5 timeline. Semafor says “moving toward releasing Fable as well, though that timeline is unclear.”
3. Broad GPT-5.6 availability. OpenAI: “we plan to make them generally available in the coming weeks.” A Sol variant on Cerebras “at up to 750 tokens per second in July” is a separate track.
4. Whether xAI, Google, or Meta are brought into the regime. WaPo’s “both” is a clue; the next step may be expansion to other frontier labs.
5. Independent benchmarks of GPT-5.6 Sol. OpenAI’s capability claims are vendor-attributed. When independent benchmarks publish, the picture may change.
6. The Annex A customer list for Mythos 5. If and when published, the list is the first concrete operational view of who the US government considers a “trusted partner.”
7. The “framework being built on the fly” materialising into a published rule. Watch for any Federal Register entry, executive order, or Commerce Department press release that names a process.

Sources

Source rule. The article’s load-bearing claims are in sources #1, #2, and #3. The article does not introduce capability comparisons, regulatory claims, customer names, or rule details not in those three sources. Any claim that would otherwise need a fourth source is omitted or attributed as “OpenAI says” / “Semafor reports” / “the Washington Post reports.”