SYSTEM CARD
The GPT-6 Astra safety record is published.
117 pages of evaluations, safeguards, deployment context, and limitations.
OpenAI’s system card is now the document to read alongside the model and API docs. This page gives the launch-day map, keeps vendor claims attributed, and calls out the limitations that matter before deployment.
CURRENT STATUS · September 7, 2026
Published September 3, 2026.
The GPT-6 Astra system card is the primary launch record for safety evaluations, cyber capability, alignment findings, monitoring, and deployment controls.
LAUNCH-DAY MAP
Five things the card changes.
The system card turns a promised document into a source that can be inspected. “Confirmed” means OpenAI published the record; it does not mean independent replication.
Capability and alignment
ConfirmedOpenAI reports stronger performance and alignment behavior than prior models across the launch evaluation record.
Read the methods and caveats before treating a vendor comparison as an independent result.
Critical cybersecurity
ConfirmedOpenAI says GPT-6 Astra is its first model to reach the Critical cybersecurity capability level.
The classification describes OpenAI’s assessment and drives stronger controls; it is not an offensive-use authorization.
Monitoring and prompt injection
ConfirmedThe launch record describes asynchronous misalignment monitoring and improved robustness to prompt injection.
The card also records monitoring limitations under adversarial conditions, so deployment context still matters.
Staged deployment
ConfirmedAstra access rolls out through controlled enterprise and defender paths before broader availability.
Permissions, tool boundaries, network scope, and human intervention remain part of the deployment decision.
Independent evidence
UnknownHowToUseAstra has not completed an independent public Astra task battery yet.
The benchmark lab is ready to record cost, latency, failures, artifacts, and review when safe access is available.
HOW TO READ IT
Keep the claim beside its condition.
Safety documents are most useful when the reader can see the task, model version, access mode, intervention policy, and failure boundary together.
- Start with scope.Confirm the exact model, snapshot, product surface, date, and evaluation population.
- Read the cyber record.OpenAI’s Critical classification and ExploitBench result describe capability under evaluation conditions; do not convert them into an operating permission.
- Read monitoring together.Stronger monitoring and prompt-injection robustness are meaningful, but adversarial monitorability limitations remain part of the record.
- Carry limitations forward.Use the card’s limitations when defining tools, approvals, logging, sandboxing, and human review for a real workflow.
WHAT IT DOES NOT PROVE
A system card is evidence, not a warranty.
It improves accountability without making the model’s tested behavior larger than the tests.
- It does not guarantee correct or safe behavior in every user workflow.
- It does not replace an application threat model, access policy, or human approval boundary.
- It does not make vendor-reported benchmark results independent evidence.
- It does not mean every ChatGPT plan, API account, region, or tool surface has identical access.
PRIMARY RECORD
Read the card at the source.
OpenAI’s system card is the primary launch document. The safety overview explains the release context, and the model guide connects the safety record to product and API behavior.
FAQ
System-card questions
Has OpenAI published the Astra system card?
Yes. OpenAI published the GPT-6 Astra system card on September 3, 2026. The card is the primary source for the model’s evaluation methods, safeguards, limitations, and deployment context.
Does the system card prove Astra is safe?
No. A system card makes tested behavior, mitigations, residual risk, and gaps easier to inspect; it cannot guarantee behavior in every prompt, tool environment, or deployment.
What is the most important limitation to understand?
The card’s safety record must be read with its deployment conditions. OpenAI reports stronger safeguards and monitoring, while also noting that monitorability can decrease under adversarial conditions relative to GPT-5.6 Sol.
ASTRA UPDATES
Keep the useful Astra updates coming
We’ll send occasional source-backed updates as access, pricing, and real-world tests change. This is an independent list, not an OpenAI waitlist.