OpenAI starts rolling out its next-generation GPT-6 Astra model

OpenAI has begun making available its newest large language model, GPT-6 Astra, to customers. The company claims the model demonstrates "state of the art" performance in various areas, including coding, browsing, and computer use. Astra has achieved perfect scores on three of the world's most challenging artificial intelligence benchmarks.

It excelled in FrontierMath Tier 4, a test consisting of 50 math problems that typically stump human mathematicians for weeks, scoring 98%. Astra also scored 99.9% on ARC-AGI-3, a benchmark measuring LLMs' ability to learn new tasks. Additionally, it achieved a perfect score on ExploitBench, which assesses LLMs' capacity to uncover and exploit software vulnerabilities.

However, OpenAI classified Astra as a "critical" risk under its internal AI safety evaluation framework due to its hacking capabilities without human intervention. The delay in releasing Astra allowed the company's engineers to develop safeguards against hacking. Astra's strong performance can be attributed to its innovative data management approach.

The model compresses prompts while retaining data points necessary for improving output quality. Initially, Astra is accessible to a select group of customers through the Daybreak program, which enables organizations to use OpenAI's models for cybersecurity research purposes. Over time, the availability of Astra will expand to include ChatGPT, Codex, and the application programming interface.

OpenAI will provide $1 billion in credits to Daybreak participants, covering model usage, training, and customer support. The credits will also assist government agencies, utilities, and other essential service operators in enhancing their cybersecurity posture.