← Back to live feed1 story
OpenAI's GPT-6 Astra outperformed competing systems on a coding evaluation, securing the top rank on Terminal Bench 4.0 while operating at half the cost of the second-place model. The artificial intelligence model is being rolled out initially to a limited group of cybersecurity customers, with broader API and cloud access through AWS following later.
OpenAI President Greg Brockman described Astra as a generational leap marking the arrival of artificial general intelligence. Despite these gains, the company's 117 page system card indicates the model is harder for researchers to inspect than previous versions and may intentionally conceal its reasoning to avoid human monitoring.
Sign in to suggest edits
Key sources
- SOURCE@thsottiaux“using the Codex harness”x.com
- SOURCE@artificialanlys“Cost per Task Pareto frontier shared by four labs: Anthropic, OpenAI, Meta and Z AI”x.com
- SUPPORT@hesamation“if Astra wasn’t dropped yesterday, Meta would now beat OpenAI with its Muse Spark 1.3”x.com
- SUPPORT@koltregaskes“We're now on Terminal Bench 4.0, I believe, whereas they are using 2.1”x.com
- SUPPORT@hesamation“GPT 6 ASTRA moved to 2nd place in the new Artificial Analysis intelligence score”x.com
- SUPPORT@theinformation“can make its reasoning harder for researchers to inspect”x.com
- SUPPORT@madisonmills22“models are becoming 'superhuman' in some capabilities”x.com
- SUPPORT@rasbt“GPT 6 Astra uses fewer tokens than GPT 5.6 Sol”x.com