New Delhi: OpenAI has officially unveiled a new AI model, dubbed GPT‑6 Astra, and OAI claims it is the most powerful and user‑friendly model yet. Astra has particularly excellent computer skills, typing, creating programming codes, Internet addicted and proficient in work-related activities such as the use of spreadsheets and presentations, according to the company.
OpenAI says that Astra achieves an accuracy rate of 98% on a challenging maths exam called FrontierMath Tier 4 and also solves math questions that experts have been unable to solve for years. It also was outstanding in a reasoning test known as ARC-AGI-3, where it is almost at the level of the human performance.
Trustworthiness and safety with the model was one of the major priorities for OpenAI. The company created a “special test” that uses an AI blunder that it encountered in the past to determine whether the model tries to do more than what it is asked to do. In the same test, earlier models sometimes overstepped and the older GPT‑5.6 Sol model went beyond its assigned task 48% of the time, Astra did not overstep at all.
However, Astra is quicker at day-to-day computer use. One test, it took approximately 47% less time to complete it, and it also scored higher than the older version.
The model’s ability to program and in cybersecurity is what stands out as well. OpenAI claims that on a metric to assess an AI’s ability to translate known software vulnerabilities into actual working exploits, Astra is 100% correct. Due to this capability, OpenAI considers Astra to fit into the “Critical” category for cybersecurity from its safety guidelines. For now, the model will not be available for hackers and malicious users to build tools for the general public; OpenAI says it will gradually open up access to trusted security researchers in the weeks ahead.
Astra is slowly being implemented. It’s currently being offered to a handful of organizations, and OpenAI states it will be widely available for all ChatGPT Plus, Pro, Business, and Enterprise users and ChatGPT developers through OpenAI’s API and Azure, Amazon Web Services (AWS).
There is a $10 price per input token (developer) and $50 per output token (developer) for pricing, and there is even a faster (more expensive) option.
