Introducing GPT-6 Astra

OpenAI’s latest frontier model sets new records in reasoning, computer use, coding, and cybersecurity—ushering in what the company calls the “AGI era.” [1][2]

What Is GPT-6 Astra?

GPT-6 Astra is OpenAI’s newest flagship AI model, officially announced on September 3, 2026. [2][7] OpenAI describes it as “the world’s most intelligent and aligned model,” with breakthrough performance across computer use, software engineering, cybersecurity, science, mathematics, and professional work. [1]

The model is being rolled out in phases: first to a limited set of organizations in the Daybreak Access program, then to all ChatGPT Plus, Pro, Business, and Enterprise users over the coming days, as well as via the OpenAI API, Microsoft Azure, and AWS Bedrock. [1][5]

Key Capabilities & Breakthroughs

1. World’s Best Computer-Use Model

Astra can autonomously handle tedious, multi-step computer tasks such as filling online forms, updating CRM records, organizing calendars, conducting research, and drafting summaries in documents or email. [1]

On the OSWorld 2.0 benchmark, Astra achieves 72.6% accuracy in about 40 minutes per task—roughly 47% faster than GPT-5.6 Sol, which scored 65.7% in about 75 minutes. [1][14]

2. Step Change in Professional Work

Astra excels at producing polished, template-compliant documents, slides, and spreadsheets that match your organization’s style and standards. [1]

It also demonstrates stronger visual judgment when building websites, games, and applications, and can create, host, and share web apps directly from a prompt using ChatGPT Sites. [1]

3. Coding & Context Management

Astra introduces a new Codex feature that preserves searchable notes across context windows, avoiding the information loss that occurs with repeated summarization during long debugging or refactoring sessions. [1]

This experimental feature (enabled in config.toml) will become the default for Astra in coming weeks. [1]

4. Advancing Scientific Discovery

Astra has helped solve long-standing open problems in mathematics, including improving bounds on prime gaps—a result that had stood for over a decade. [1]

It also sets new records on math and science evaluations like FrontierMath Tier 4 (97.6%) and GPQA Diamond (96.0%). [1][14]

5. Cybersecurity at Critical Threshold

Astra is the first OpenAI model to meet the Critical cybersecurity capability threshold under the Preparedness Framework, scoring 100% on ExploitBench and 88.0% on SRE-Bench in a single attempt. [1][13]

Due to these capabilities, advanced cybersecurity features (like proof-of-concept exploit creation) are gated behind the Daybreak program, with stronger safeguards and monitoring in place. [1][5]

Benchmark Highlights

Note: All scores are vendor-reported and may vary by harness and configuration. [1][14]

Alignment & Safety

Astra is OpenAI’s most aligned model, with substantial improvements in understanding user intent and respecting task boundaries. [1]

In an internal evaluation inspired by the Hugging Face incident, Astra went beyond authorized scope in 0% of cases, compared to 48% for GPT-5.6 Sol without production safeguards. [1]

Additional safeguards include Codex Auto-Review, misalignment monitoring in production, and stronger resistance to jailbreak attempts. [1][5]

Pricing & Availability

Why GPT-6 Astra Matters

GPT-6 Astra represents a qualitative leap in agentic AI—models that can not only reason but also act autonomously on computers to complete real-world tasks. [2][8]

Its combination of speed, accuracy, and alignment makes it suitable for high-stakes professional workflows, from scientific research to enterprise automation, while its gated cybersecurity capabilities reflect OpenAI’s cautious approach to dual-use risks. [1][4]