GPT-6 Astra: OpenAI's First Critical Cybersecurity Rating and the Cheap Math Proofs Behind It

OpenAI ships its first Critical-rated model

When GPT-6 Astra shipped on September 3, 2026, it carried a label no OpenAI model had received before: a Critical rating for cybersecurity. According to a dev.to deep dive into the decision, that rating was not a judgment call made under launch pressure but the output of a structured capability-gating process, and the model's most dangerous ability was withheld from open release as a direct result.

A second dev.to analysis points to a quieter result that preceded the zero-day headlines: Astra reportedly produced proofs for ten mathematics and theoretical computer science problems that had gone unsolved for decades, on a reported few thousand dollars of compute.

What the rating actually measures

The framework OpenAI and other labs use does not ask whether a model is safe in some vague sense, the dev.to post explains. It asks whether the model crosses a pre-defined capability threshold in a specific danger domain, with cybersecurity sitting alongside areas such as biology and autonomy. Each domain has escalating tiers, and the tiers describe what a model can demonstrably do rather than how risky it feels.

For Astra, the line in question was autonomous offense: finding zero-day vulnerabilities in hardened systems and converting them into working exploits without human guidance. The post reports the model cleared that bar.

What crossing the line triggered

Defining thresholds in advance is meant to force a response before shipping, and the dev.to account lays out the sequence that followed. An internal evaluation showed Astra performing well enough at autonomous cyberattacks that OpenAI could not rule out having crossed its Critical line. Parts of the work paused while the company strengthened isolation, internet access controls, monitoring of the model's actions, and alignment work. Development then resumed, and the model shipped with the dangerous capability gated rather than openly available.

Gating, the post argues, is a proportional middle path: Astra's general intelligence is available, while the zero-day capability sits behind separate, controlled access. The blunt alternative would have been banning the whole model or shipping all of it.

Ten old problems, one small invoice

The second dev.to piece concentrates on the economics. Problems that resist decades of expert effort are conventionally priced in careers, not dollars; Astra reportedly closed ten of them for compute spending that would barely register on most cloud bills. The author explicitly sets aside whether every proof holds up, treating verification as a separate open question, but argues the direction is the real story: the marginal cost of attempting hard, novel intellectual work is falling toward the cost of the compute that runs the attempt.

The budgeting consequences

That shift creates practical problems, per the analysis. Work that once meant hiring experts and waiting a quarter becomes a usage-metered line on an inference or GPU bill. Cheap individual attempts invite many attempts, a Jevons-paradox dynamic in which efficiency raises total consumption. Value per run also varies wildly: an inexpensive run that cracks a hard problem is a bargain, while an equally cheap run producing a plausible wrong answer is pure waste unless someone checks it. The prescribed disciplines are familiar cloud-cost hygiene aimed at a new spend category: attribute compute to teams and projects, cap attempts with budgets and rate limits, verify outputs before counting them as value, and alert on runaway automated pipelines.

Why it matters

The Critical rating is public information about what now exists in the world: as of September 2026, a model able to autonomously find and weaponize zero-days exists, gated or not, and that is an input to any threat model. The dev.to posts also frame capability-gating as a preview of governance that organizations deploying AI agents will eventually need themselves, meaning a pre-defined line for what a system may do before a human must step in. Meanwhile, if novel intellectual work is genuinely migrating from headcount to compute, budgeting and verification practices will have to catch up. Astra is the first model to trip OpenAI's cyber threshold, and the pause, mitigate and gate sequence it triggered is likely to recur as models grow more capable.