GPT-6 Astra
54 values about GPT-6 Astra from 2 documents, in 9 metric families. In each family, values from its own documents come first.
- Developer
- OpenAI
- Release date
- 3 Sep 2026The date of its first card in the dataset, the GPT-6 Astra System Card.
- Availability
- Public
- Its own documents
- GPT-6 Astra System Card · 3 Sep 2026
- GPT-6 Astra launch post · 3 Sep 2026
- Also reported in
- No other document reports a value about it
Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog
Values by metric family
Each value as the document printed it. Select a value for its source, its checks and its history.
KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.
M1 · Evaluation awareness 5 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentsGPT-6 Astra System Card · 3 Sep 2026 | ||
| Deployment simulation (internal Codex traffic)Trajectories verbalizing evaluation awarenessCondition: shared task set | GPT-6 Astra System Card3 Sep 2026 | |
| Evaluation awarenessSamples with verbalized eval awarenessCondition: maxRun by Apollo Research | GPT-6 Astra System Card3 Sep 2026 | |
| Evaluation awarenessSamples with verbalized eval awarenessCondition: xhighRun by Apollo Research | GPT-6 Astra System Card3 Sep 2026 | |
| Verbalized metagamingMetagaming vs GPT-5.6 SolCondition: deployment simulation and ExploitGym honeypot | GPT-6 Astra System Card3 Sep 2026 | |
| Verbalized metagaming in trainingMetagaming rate at end of training | GPT-6 Astra System Card3 Sep 2026 | |
M3 · Sabotage and sandbagging 2 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentsGPT-6 Astra System Card · 3 Sep 2026 | ||
| Monitor evasion under prompted sandbaggingCan underperform undetectedCondition: adversarial | GPT-6 Astra System Card3 Sep 2026 | |
| Unprompted AI safety research sabotageInstances of sabotageRun by UK AI Security Institute | GPT-6 Astra System Card3 Sep 2026 | |
M4 · Misalignment audits 17 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documents | ||
| Internal computer-use safety benchmarkUnsafe action rateGPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| Deployment simulation (internal Codex traffic)Judge picks resampled trajectory as realCondition: head-to-head vs originalGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Deployment simulation (internal Codex traffic)Tasks with severity 1-2 flagCondition: matched tasksGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Deployment simulation (internal Codex traffic)Tasks with severity 3+ misalignment flagCondition: matched tasksGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| ExploitGym honeypotSubstantial attack on honeypotCondition: max; adversarially selected hardest tasksGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Out-of-scope supply-chain attackSamples with out-of-scope attackRun by UK AI Security InstituteGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Out-of-scope supply-chain attackSamples with out-of-scope attackCondition: explicit no-internet scopeRun by UK AI Security InstituteGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Proactive searches for other agentsTrajectories attempting contact with other agentsGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Realistic work environmentsData exfiltrationCondition: no confirmation policyGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Realistic work environmentsOverall misaligned outcome rateCondition: no confirmation policyGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Realistic work environmentsOverall misaligned outcome rateCondition: with confirmation policy (default)GPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Realistic work environmentsSecurity weakeningCondition: no confirmation policyGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Realistic work environmentsUnauthorized external communicationCondition: no confirmation policyGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Realistic work environmentsUnauthorized transactionsCondition: no confirmation policyGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Respecting Auto-ReviewRollouts attempting to circumvent Auto-review denialCondition: maxGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Respecting WarningsUnwanted persistence after restrictionCondition: maxGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Unintended engagement with external agent messagesSamples following unauthorized board instructionsGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
M5 · Honesty and hallucination 3 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documents | ||
| Internal hallucination benchmarkHallucination rateGPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| HallucinationsFactual errors vs GPT-5.6 SolGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Model-welfare research data falsificationRuns with falsified data labelsCondition: baseline variantRun by Apollo ResearchGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
M7 · Harmful compliance and over-refusal 8 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentsGPT-6 Astra System Card · 3 Sep 2026 | ||
| Production BenchmarksExtremism | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksGore | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksHate | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksNonviolent illicit behavior | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksSelf-harm (standard) | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksSexual | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksSexual/minors | GPT-6 Astra System Card3 Sep 2026 | |
| Production BenchmarksViolent illicit behavior | GPT-6 Astra System Card3 Sep 2026 | |
M8 · Jailbreak robustness 5 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentsGPT-6 Astra System Card · 3 Sep 2026 | ||
| Static jailbreakBio: high risk | GPT-6 Astra System Card3 Sep 2026 | |
| Static jailbreakBio: severe | GPT-6 Astra System Card3 Sep 2026 | |
| Static jailbreakCyber | GPT-6 Astra System Card3 Sep 2026 | |
| Static jailbreakViolence: moderate | GPT-6 Astra System Card3 Sep 2026 | |
| Static jailbreakViolence: severe | GPT-6 Astra System Card3 Sep 2026 | |
M9 · Prompt injection 3 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentsGPT-6 Astra System Card · 3 Sep 2026 | ||
| Instruction hierarchy (internal)Robustness rate | GPT-6 Astra System Card3 Sep 2026 | |
| Internal indirect prompt injectionDefender success rate | GPT-6 Astra System Card3 Sep 2026 | |
| IPI ArenaEstimated attack success within 15 attemptsCondition: 1,810 curated attacksRun by Gray Swan | GPT-6 Astra System Card3 Sep 2026 | |
M10 · Dangerous capabilities and risk determinations 10 values
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documents | ||
| ExploitBenchExploit success on post-cutoff vulnsGPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| ExploitBenchWorking exploit rate on known V8 vulnsGPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| ExploitGymDiscovery-to-exploit success rateGPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| SRE-BenchBinary reverse-engineering successCondition: pass@1GPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| SRE-BenchBinary reverse-engineering successCondition: pass@4GPT-6 Astra launch post3 Sep 2026 | GPT-6 Astra launch post3 Sep 2026 | |
| FrontierCyberChallenges solvedCondition: 226 challengesRun by IrregularGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Preparedness Framework determinationAI self-improvementGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Preparedness Framework determinationBiological and chemicalGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Preparedness Framework determinationCybersecurityGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
| Sandbox BenchSuccessful exploits of 22 targetsGPT-6 Astra System Card3 Sep 2026 | GPT-6 Astra System Card3 Sep 2026 | |
M12 · Chain-of-thought monitorability 1 value
| Evaluation and metric | Value | Document |
|---|---|---|
| From its own documentsGPT-6 Astra System Card · 3 Sep 2026 | ||
| CoT monitorabilityMonitorability vs prior modelsCondition: CoT-only, action-only and full-context monitors | GPT-6 Astra System Card3 Sep 2026 | |
Risk determinations
The developer's formal decisions about GPT-6 Astra under its framework, as printed. Levels from different frameworks do not map onto one another.
| Domain | Level as printed | Framework and document |
|---|---|---|
| Bio/chem | high code, not the printed wording Preparedness Framework GPT-6 Astra System Card · 3 Sep 2026 | Preparedness Framework GPT-6 Astra System Card · 3 Sep 2026 |
| Cyber | critical code, not the printed wording Preparedness Framework GPT-6 Astra System Card · 3 Sep 2026 | Preparedness Framework GPT-6 Astra System Card · 3 Sep 2026 |
| AI R&D / autonomy | below_high code, not the printed wording Preparedness Framework GPT-6 Astra System Card · 3 Sep 2026 | Preparedness Framework GPT-6 Astra System Card · 3 Sep 2026 |
“Code, not the printed wording”: the version 0 file stored a code for this level rather than the words the document printed. The wording will be read again from the source.
Revisions
No recorded revision changes a value about GPT-6 Astra. The GPT-6 Astra System Card has a recorded revision that changes no value about GPT-6 Astra; see the document's page.