Report: GPT-5.6 Preview opens customer-by-customer during federal review

• According to Stephanie Palazzolo's report and The Information's story, OpenAI is holding GPT-5.6 to a limited preview because the federal government asked it to stagger the release.
• Kolt Regas's summary of the report and Yuchen Jin quoting the key line both center the unusual part: access during the preview is reportedly being approved customer by customer.
• Pre-release traces already showed gpt-5.6-preview, gpt-5-6-thinking, and gpt-5-6-auto in ChatGPT and Codex surfaces, with TestingCatalog's route screenshot, Kolt Regas's enterprise sighting, and scaling01's auto model ID pointing to rollout plumbing that existed before the report.
• thdxr's read argues this could still end in a normal public launch after review, while Levie's thread frames the bigger change as a de facto review gate for frontier releases.

You can read The Information's report, check Anthropic's still-live Fable and Mythos access note, and see that the rollout plumbing was already peeking through in a Codex repo screenshot and a ChatGPT route leak.

The core claim is narrow and concrete. OpenAI did not cancel GPT-5.6. It reportedly shifted the model into a limited preview, with approvals handled customer by customer during the review period, and staff were told a broader release could follow a couple of weeks later if things go well, per kimmonismus posting The Information excerpt.

That timeline matters because much of the reaction online jumped straight to an indefinite freeze. thdxr's thread is one of the few posts that keeps the original framing intact: restricted access during an unreviewed window, then a possible normal release.

The report ties OpenAI's handling of GPT-5.6 to what already happened at Anthropic. kimmonismus summarizing The Information says Anthropic had already limited Mythos to select partners, and Wes Roth citing Sam McAllister later said Anthropic was serving exactly zero traffic to Fable 5, with its brief appearance in Claude possibly just a UI bug.

That makes GPT-5.6 look less like a one-off delay and more like a new release pattern for models that trip whatever security threshold Washington is now using. The useful external corroboration here is Anthropic's own Fable and Mythos access note, which confirms that access to those models was being handled separately from a normal public rollout.

The policy wrinkle is that the public framing around review has not sounded like formal licensing. kimmonismus notes that the administration's executive-order process was supposed to be voluntary, while an Axios excerpt screenshot says Commerce Secretary Howard Lutnick wanted GPT-5.6 tested and approved across relevant parts of government because the model allegedly has "Mythos-like" capability.

That leaves two claims on the table at once:

• The formal policy is described as review, not preclearance, according to kimmonismus's summary of the order language.
• The operational reality described in the reporting is access approval customer by customer, according to Stephanie Palazzolo's report and Axios's follow-up excerpt.

A lot of the engineer reaction sits exactly in that gap. Levie calls it de facto AI regulation, while Ethan Mollick argues the missing piece is public detail on what risks are actually triggering intervention.

Weeks before the report, GPT-5.6 was already leaking through product surfaces:

• admin/model-access/gpt-5.6-preview, in TestingCatalog's ChatGPT route leak and haider1's earlier route sighting
• gpt-5-6-thinking, visible to some enterprise users, in Kolt Regas's enterprise sighting
• gpt-5-6-auto, in scaling01's model ID leak
• Codex tests that explicitly kept GPT-5.6 from becoming the default model, in aibuilderclub_ reading the test and the repo screenshot

The Codex detail is the weirdest one. aibuilderclub_ says tests were added so GPT-5.6 would not become the default even if the server returned it, which looks a lot more like controlled rollout machinery than a stray placeholder.

The direct consequence of a customer-by-customer preview is simple: most people cannot benchmark or integrate the model on day one because they do not have the model. That was already visible in the pre-report leaks, where enterprise sightings and route strings appeared without a general launch, per Kolt Regas's enterprise sighting and Kolt Regas's preview-only note.

The secondary consequence is that a lot of market knowledge stays bottlenecked. If the first wave of access is a small partner set, then early evals, jailbreak reports, coding regressions, and pricing behavior surface unevenly, if at all. Ethan Mollick's thread makes the same point from a different angle: firms are being asked to prepare for risks they are not being told much about.

The most concrete reason not to overstate this story is buried in the original wording. The Information excerpt screenshot says Altman hoped for a more general release a couple of weeks later if all went well, and thdxr reads the episode as a review period, not a permanent access regime.

That does not settle the bigger policy fight. It does add one important fact that was missing from a lot of the reaction posts: the reporting describes a staged release path, not a canceled model.