OpenAI deploys GPT-6 Astra, its most powerful AI model, capable of using a single computer

OpenAI began deploying GPT-6 Astra on Thursday, presented as its most powerful artificial intelligence (AI) model, capable according to the San Francisco company of using a computer on its own and distributed with a device that can stop it if it leaves its frame. This generation leap, a year after GPT-5, appears after a series of cyberattacks carried out this summer by AI in the testing phase, including tools from OpenAI and its rival Anthropic, which illustrated the risks posed by cutting-edge models.

GPT-6 is initially reserved for “a limited number of organizations”, before being opened “in the coming days” to paying ChatGPT subscribers and developers. In a demo video, the company shows its voice-controlled model publish an online sales ad on its own. The model can find accommodation in ten minutes, compared to six hours for a human tester, according to test results posted by OpenAI.

Skip the ad

Highest risk level

GPT-6 is OpenAI's first model to reach the highest risk level of its own cybersecurity assessment grid, capable of hacking into securely protected computer systems on its own. Its most sensitive functions are reserved for cyber defense specialists and automatic monitoring can interrupt user tasks, “including work unrelated to cybersecurity”.

For its part, Anthropic has been reserving its Mythos model for select partners since April, also capable of carrying out hacks alone, of which a restricted version, Fable, has been open to the public since June. OpenAI argues that GPT-6 accomplishes every task for less than its predecessor, but still at some of the highest prices on the market, like Anthropic.

Asked about the competition from low-cost open models, particularly from China, which could threaten the economic model of the American champions, the president of OpenAI, Greg Brockman, appeared confident in front of journalists: “what matters is the price per task” achieved, and not the price of the credit for using the computing power (token). Greg Brockman felt, personally, that it was “not unreasonable” to consider that technology was entering the era of AGI, the English acronym for artificial general intelligence, this Silicon Valley holy grail designating the stage where AI would be comparable to the performance of human intelligence.