OpenAI has launched Astra, a model it describes as "a new frontier on computer and browser use" and its "best model for software engineering to date," TechCrunch reports. It went first to customers of Daybreak, OpenAI's cybersecurity programme, on 3 September, with the Pro, Plus, Business and Enterprise plans and the API getting access within the following week.
What it is built for
In OpenAI's own benchmarks, Astra scores higher than existing models, including OpenAI's Sol and Anthropic's Fable, on finding bugs, executing terminal tasks and answering questions about codebases. OpenAI also says Astra's ability to "identify and develop zero-day exploits" can help defenders find and patch weaknesses before attackers do. The company did not publish numerical scores in the coverage.
OpenAI president Greg Brockman called it the company's "most intelligent and, also very importantly, our most aligned model yet."
The catch
Astra uses a reasoning technique called "opaque recurrence," which can hide the step-by-step chain of thought that researchers read to check how a model reaches its decisions, TechCrunch reports. OpenAI downplays how much the model uses it.
"As model capabilities are increasing, monitorability is getting more challenging," said OpenAI chief scientist Jakub Pachocki, according to TechCrunch.
