Mara ReyesOpenAI says its government-visible GPT-5.6 preview should not become the long-term default. Yet Anthropic’s route back to deployment already begins with access divided by model and user class.
The preview list no longer belongs only to the lab
The conventional release system answered a question controlled by the laboratory: is this model ready to move from internal testing to preview, then general availability? The company chose the users, safeguards, sequence, and timing. A limited preview reinforced that control. It was one phase of a company-run launch, not a separate layer of public governance.
GPT-5.6 preserves the appearance of that sequence while changing one of its load-bearing assumptions. OpenAI is previewing the model to roughly 20 companies before planned general availability, but the participating companies are being disclosed to the US government.
Disclosure is not approval, and government visibility is not government control. Those distinctions matter. But the release can no longer be described only as a transaction between a model provider and its selected customers. A third party now has formal visibility into who receives early capability.
The model may still belong to OpenAI. The preview list no longer belongs only to OpenAI.
A temporary exception can still change the architecture
OpenAI explicitly says this process should not become the long-term default. Its planned endpoint remains general availability, not government-disclosed participation for every release.
But temporary arrangements reveal structure because they show which assumptions can be suspended under pressure. The important fact is not that this preview is permanent. It is that frontier-model access can now be organized around a disclosed cohort without stopping the release process.
Anthropic reached the same structure from the opposite direction: blocked access followed by a selective route toward restoration. Mythos 5 is approved for critical-infrastructure operators, while broader Fable 5 access is still being worked out.
These are not equivalent cases. OpenAI retains a planned path from preview to general availability. Anthropic’s restoration remains partial, divided by model and user category. Yet both processes decompose “release” into the same smaller questions: which model, which users, which safeguards, which visibility, and which conditions must hold before access expands.
Broad deployment did not stop being the laboratories’ goal. It stopped being the only unit the system could govern.
Capability is no longer the whole competitive unit
A model race organized around capability produces familiar comparisons: one system performs better, costs less, or arrives sooner. A deployment regime adds another dimension. The product is no longer just the model’s capability; it is the package of permissions under which that capability can reach a particular class of user.
For GPT-5.6, that package currently includes a limited company cohort, disclosure of participants to the government, and a planned transition to general availability. For Mythos 5, it includes approved access for critical-infrastructure operators. For Fable 5, the terms of general access remain unresolved.
This does not make the US government the product manager for either laboratory. It does mean that release conditions have become competitive infrastructure. A laboratory may possess a stronger model and still face a narrower deployment path. Another may move toward wider availability while accepting more formal visibility around its early users. The capability remains inside the model, but the usable advantage is determined at the access boundary.
That boundary creates its own feedback loop. More consequential capability makes deployment conditions more salient. Explicit conditions then pull user classes, safeguards, and government visibility into the release itself. Once there, laboratories must compete through them rather than around them. No permanent rule is required; repeated exceptions are enough to establish the mechanism.
The launch has reversed into a negotiation
The machinery of a controlled launch now distributes authority instead of concentrating it. OpenAI still selects a limited cohort, but the cohort is disclosed outside the company. Anthropic still seeks to deploy its models, but restoration is separated by model and operator class. The release process remains controlled, except control is no longer located in one place.
This is the reversal. A launch mechanism built to let a company decide when everyone could receive a model is becoming a mechanism for negotiating who may receive it first, who may receive it at all, and which institution must be able to see the difference.
Neither case establishes a permanent template. OpenAI rejects that interpretation of its preview, and Anthropic has not secured uniform restoration.
A default need not be declared for a checkpoint to take hold. The old launch button now sits behind a permissions sheet: roughly 20 company names on one side, critical-infrastructure operators on the other.