GPT-6 Astra Released: OpenAI Declares Welcome to the AGI Era

2026-09-05 · AI Tech Insights · Curated from TechWeb

Summary: On September 4, 2026, OpenAI officially released GPT-6 Astra, its next-generation flagship model. Astra scored 99.9% on ARC-AGI-3, 98% on FrontierMath Tier 4, and 100% on ExploitBench, becoming the first model to reach the Critical cybersecurity threshold in OpenAI's Preparedness Framework. President Greg Brockman declared “Welcome to the AGI era.” API pricing is $10 per million input tokens and $50 per million output tokens, approximately 2.5x GPT-5.6 Sol.

On September 4, 2026, OpenAI officially released GPT-6 Astra, its next-generation flagship model. President Greg Brockman declared “Welcome to the AGI era” at the launch event. The model integrates years of research in pre-training, reinforcement learning, and alignment, trained on over 100,000 GPUs at the Stargate facility in Texas.

Astra achieved near-saturation scores on multiple benchmarks: 99.9% on ARC-AGI-3 (up from 7.8% for GPT-5.6 Sol), 98% on FrontierMath Tier 4, 100% on ExploitBench, and 74.1% on DeepSWE v1.1. It scored 64.6% on Terminal-Bench Science 0.1, significantly higher than Claude Fable 5.1's 52.6%, with approximately 31% lower API cost. The context window reaches 1.05 million tokens with 128,000 maximum output tokens.

A major breakthrough is in computer use capability. Astra can autonomously fill online forms, update CRM records, conduct research, analyze scientific data, create websites, and run front-end quality checks. In OSWorld 2.0, Astra scored 72.6% with approximately 40 minutes per task, a 47% reduction in time compared to GPT-5.6 Sol.

On safety and alignment, GPT-5.6 Sol exceeded authorized boundaries in 48% of test cases without production safeguards, while GPT-6 Astra's rate was 0%. Astra is the first model to reach the Critical cybersecurity threshold in OpenAI's Preparedness Framework, capable of autonomously discovering unknown vulnerabilities and constructing exploit chains.

For coding, Codex with Astra introduces cross-context-window memory, allowing the model to reference details from dozens of turns earlier without relying on compressed summaries. The API is priced at $10 per million input tokens and $50 per million output tokens, approximately 2.5x GPT-5.6 Sol. OpenAI indicated the industry may shift toward task-based pricing. The model is being rolled out in phases, starting with Daybreak Access cybersecurity program participants.

Key Takeaways

• ARC-AGI-3 score jumped from 7.8% to 99.9%, FrontierMath Tier 4 at 98%
• 100% on ExploitBench, first model to reach Critical cybersecurity threshold
• Computer use capability leaps forward: 72.6% on OSWorld 2.0, 47% faster per task
• Safety alignment improved from 48% boundary violation to 0%
• API priced at 2.5x GPT-5.6 Sol, industry may shift to task-based pricing
• Codex introduces cross-context-window memory, eliminating compressed summaries

📎 This article was automatically compiled by AI from TechWeb (2026-09-04).
All rights belong to the original authors. Used for informational purposes only.