OpenAI's GPT-6 Astra is here — and it's the first AI to trigger a 'Critical cyber' warning

OpenAI launched GPT-6 Astra on September 3, 2026, making it available first to a limited group of organizations before rolling out to ChatGPT Plus, Pro, Business, and Enterprise subscribers over the coming days. The model scores 99.9% on ARC-AGI-3 — a leading measure of general reasoning — and 100% on ExploitBench, a test of software vulnerability exploitation. That last number is why Astra is also the first model to cross OpenAI's newly defined "Critical cyber threshold," triggering restrictions that no previous model required.

The claim

Astra is built for autonomous work: it can control a computer or browser, run multi-step tasks in parallel, and generate 3D models, presentations, and code without constant prompting. OpenAI says it follows initial instructions while analyzing visual information as it goes — closer to a delegated assistant than a chatbot. In benchmarks it also beats the previous flagship, GPT-5.6 Sol, completing equivalent tasks roughly 47% faster, per DataCamp.

The cybersecurity score is the headline complication. Astra hit 100% on ExploitBench, which measures a model's ability to find and exploit real software flaws. OpenAI says consumer and API versions carry safety filters and abuse monitoring to prevent dangerous use, but full offensive cyber capabilities are restricted to the Daybreak program — a defense-only channel unavailable to general users. UK regulators were not part of the pre-launch briefing; OpenAI went through White House vetting but no FCA or ICO pre-approval has been reported, per CNBC.

Availability and cost

ChatGPT Plus subscribers in the US and UK get Astra access as part of their existing plan — no extra charge on top of the current monthly fee. API access opens alongside the consumer rollout, priced at $10 per million input tokens and $50 per million output tokens. That is 2.5 times the promotional rate of GPT-5.6 Sol, though OpenAI argues the speed gain offsets the cost for professional workloads.

EU users face an extra constraint: OpenAI's EU data residency option adds a 10% pricing uplift and blocks access to Fast mode entirely — a gap that will matter for European enterprise customers comparing costs against US counterparts.

Safety in practice

The UK's AI Safety Institute flagged an evasion risk under adversarial prompting during pre-release evaluation, according to Pillitteri. OpenAI acknowledges the finding and says ongoing monitoring is built into every deployment tier. For most users — writing, coding, research, document work — Astra functions like a significantly more capable version of what came before. The cyber restrictions are real, but they apply to a capability set that general users won't encounter in day-to-day use.