GPT-6 Astra release notes / what's new

GPT-6 Astra enables direct computer use for professional workflows

GPT-6 Astra is designed to operate within existing business applications and workflows without requiring custom APIs or data preparation. By utilizing computer use capabilities, Astra can interact with software as a human does, allowing it to perform tasks across various applications from day one.

Key professional applications include:

  • Financial Modeling: In challenges from the Financial Modeling World Cup (based on the 2023 Microsoft Excel World Championship), GPT-6 Astra completed tasks approximately four times faster than the winning human competitor.
  • Software Engineering: OpenAI's internal engineering team used Astra to resolve a memory-allocation bottleneck in a test environment, achieving 25× lower turn latency with a roughly 30% increase in peak memory use.
  • Content Creation: The model can adhere to company-specific voice, templates, and design standards, such as creating slide decks that match a business's specific layout and tone.

Performance and cost-efficiency benchmarks

GPT-6 Astra is engineered to complete tasks using fewer tokens and fewer retries, reducing the cost per task. OpenAI reports that Astra occupies the majority of the cost-efficiency frontier for professional work and coding evaluations.

Terminal Bench 4.0 Results

On Terminal Bench 4.0, which evaluates agents on complex terminal-based tasks including system configuration and data analysis, GPT-6 Astra achieved a score of 57.9%. This outperforms previous models:

  • GPT-5.6 Sol: 37.3% (Astra is approximately 9% lower in estimated API cost per task).
  • Claude Fable 5.1: 55.8% (Astra is approximately 63% lower in estimated API cost per task).

Pricing

API pricing for GPT-6 Astra starts at $10 per million input tokens and $50 per million output tokens.

Safety, alignment, and enterprise control

GPT-6 Astra is OpenAI's most aligned model to date, featuring increased adherence to human intent and authorization to ensure safe operation within business systems.

Computer Use Safety Benchmark

In an internal computer use safety benchmark testing scenarios like data deletion or the exposure of confidential information, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1.

Enterprise Administrative Controls

To manage deployment, OpenAI has introduced new enterprise admin controls that allow organizations to:

  • Restrict access to approved websites and desktop applications.
  • Manage uploads and downloads.
  • Control browsing history.
  • Implement confirmation policies requiring approval before consequential actions.
  • Perform automated reviews of potentially unsafe or unauthorized tool calls.

Cybersecurity and Preparedness

GPT-6 Astra is the first model to reach the "Critical cybersecurity capability threshold" under OpenAI's Preparedness Framework. In response, OpenAI has strengthened protections against misuse and unauthorized actions, including training the model to respect security boundaries and deploying automated checks to block harmful responses.

Availability and Integrations

GPT-6 Astra is available in ChatGPT Work, Codex, and via the API. To expand access, OpenAI is launching new enterprise plugins in ChatGPT Desktop for Oracle Analytics, Power BI (a Microsoft Fabric service), Navan, and Avalara.

Sources