Overview
- OpenAI released GPT‑6 Astra on Thursday and began a phased rollout that starts with vetted Daybreak partners and will expand to ChatGPT paid tiers and API partners in the coming days.
- In internal testing Astra scored 100% on ExploitBench and discovered and chained two previously unknown zero‑day vulnerabilities that OpenAI is disclosing to affected maintainers.
- OpenAI says Astra can act inside software and browsers to perform complex multi‑step professional tasks such as PCB layout, 3D scenes, coding and tax‑form drafting rather than only offering recommendations.
- Because Astra is harder to monitor under adversarial tests, OpenAI is prioritizing chain‑of‑thought and activation monitoring, stronger sandboxing, identity and legal gates, and keeping its most advanced cyber features restricted to vetted defenders.
- The release includes large technical limits and pricing that matter for adoption — a roughly 1,050,000‑token context window, up to 128,000 output tokens, and API rates starting at $10 per million input tokens and $50 per million output tokens — and it has renewed calls from security experts and policymakers for standardized pre‑release testing and forensic reviews.