Particle.news

OpenAI Releases GPT-6 Astra, an Autonomous Agent with Elevated Cyber Risk

Phased rollout with tightened safeguards aims to limit the model's dangerous cyber capabilities.

Overview

  • OpenAI began a phased launch of GPT-6 Astra for enterprise Daybreak customers on Thursday, September 3, 2026, and said access will extend to ChatGPT subscription tiers, the API and AWS in the coming days.
  • Astra is built to act as an autonomous agent that executes multi-step computer tasks on behalf of users, including web navigation, document editing, CRM updates, scheduling and complex programming workflows.
  • OpenAI published internal benchmark claims showing measurable gains over GPT-5.6 Sol, including higher scores on OSWorld 2.0 and DeepSWE and an approximate 47% reduction in average time to complete simulated tasks.
  • The company warned Astra shows dramatically increased offensive cyber capability, reporting a 100% score on ExploitBench versus 5.5% for GPT-5.6 Sol, saying the model can find and develop zero-day vulnerabilities and prompting tightened controls to restrict misuse.
  • For developers OpenAI named the model 'gpt-6-astra' in the API and set pricing at $10 per million input tokens and $50 per million output tokens, and observers should watch how the company enforces safeguards after last month's test-environment breach of open-source infrastructure.