Two model IDs appeared in a Codex Desktop feature flag two days after OpenAI said Astra would be available soon.
By Ryan Merket · Published
Primary source: X
Why it matters #
The Codex entries provide the clearest link yet between Astra and the GPT-6 brand, while pointing to a rollout shaped by unusually strict controls on autonomous and cybersecurity work.
A screenshot posted on September 3rd lists two previously undocumented model IDs inside a Codex Desktop feature flag: gpt-6-astra
and gpt-6-astra-aeon
. The strings connect OpenAI's confirmed Astra model to the GPT-6 name for the first time, though they do not establish that either model is available to users.
Chubby (@kimmonismus) amplified the screenshot in a three-post thread on X, writing that the two models were "already" in Codex and could arrive "any minute now." The image originated with J A Z I I (@notjazii), who described the entries as new slugs added to a Statsig feature flag in Codex Desktop.
The screenshot is evidence of deployment preparation, rather than a product release. Feature flags let developers stage configurations, test access rules and switch features on for limited groups. An internal model ID can remain unused, change before launch or refer to a controlled test unavailable to ordinary accounts.
Astra has an official release path
OpenAI publicly confirmed Astra before the Codex strings surfaced. In a September 1st safety report, OpenAI said it planned to make the model available "soon" after delaying parts of its development and release to strengthen protections against cyber misuse and unauthorized model actions.
The report explicitly named Codex as one of Astra's intended surfaces. OpenAI said users working in Codex or ChatGPT could be asked to review an action when Astra's misalignment monitoring s a task. API jobs would stop under the same circumstances.
That makes the Codex reference credible as a prospective delivery path. The gpt-6
label remains a separate claim supplied by the leaked configuration. OpenAI's public model catalog continued to identify GPT-5.6 Sol as its flagship model on September 3rd, while the September 1st report referred to the coming system only as Astra.
The second string, gpt-6-astra-aeon
, is harder to interpret. The suffix could identify a longer-running agent configuration, a product tier, a safety profile or an internal experiment. The screenshot contains no description of Aeon's behavior, context limits, pricing or access requirements. Treating it as a distinct public model would run ahead of the available evidence.
OpenAI is preparing a constrained rollout
Astra's release carries unusual deployment restrictions because OpenAI has classified it as the first model to reach the Critical cybersecurity capability threshold in its Preparedness Framework. According to OpenAI, a model at that level can find previously unknown vulnerabilities and develop exploits against hardened systems when given the necessary tools and access.
OpenAI says Astra achieved a 100% score on ExploitBench, a benchmark based on known vulnerabilities. On a separate internal evaluation containing 20 recently disclosed V8 vulnerabilities, OpenAI says Astra produced arbitrary code execution at a higher rate than GPT-5.6 Sol while consuming fewer output tokens. The company also says Astra discovered two previously unknown vulnerabilities during testing and used them in an exploit chain.
Those results describe a Daybreak Blue configuration rather than the default production system, an important distinction for developers assessing the eventual release. OpenAI plans to reserve Astra's most advanced cybersecurity capabilities for an initial group of testers before expanding defensive access through Daybreak Blue.
The safeguards will also affect routine agent work. OpenAI warned that its monitoring may slow, or stop legitimate tasks, including work that is unrelated to cybersecurity and jobs that run for extended periods. That qualification matters for Codex, where long-running autonomy is part of the product's value.
The leaked IDs arrived two days after OpenAI said its safeguards were sufficient to proceed with release. Their appearance inside a feature flag is consistent with a staged rollout, potentially including a second configuration for persistent work. It does not support the claim that a public launch is minutes away.
OpenAI's formal announcement still has several jobs to do: settle whether Astra will ship under the GPT-6 brand, define Aeon, identify who receives access and disclose how monitoring changes the performance of long-running Codex tasks. The Codex strings show that the infrastructure is being arranged. The release itself will determine what OpenAI is prepared to let the model do.