OpenAI launches GPT-6.1 Sol, a cheaper and safer AI model, amid security pressure

The creator of ChatGPT, OpenAI, launched on Tuesday a new intermediate artificial intelligence (AI) model presented as almost as efficient as its flagship model for five times less expensive, the day after the abandonment of a new version of the latter, deemed unsafe. Called GPT-6.1 Sol, this model is one step below GPT-6 Astra, OpenAI's most powerful model, launched on September 3. Its release, during the group's annual developer conference, is part of the price battle against its rival Anthropic, before their highly anticipated IPOs.

The company told AFP on Monday that it would not launch GPT-6.1 Astra, the new version of its flagship model, believing that it too often went beyond its scope. According to its own tests, GPT-6.1 Sol respects the instructions better than the version which preceded it. It attempts to circumvent an explicit ban almost three times less often, but remains less disciplined than GPT-6 Astra. GPT-6.1 Sol is reserved for paying subscribers of ChatGPT, in its tools dedicated to work and programming, as well as developers.

Skip the ad

Calls to slow down

The serial revelations, since July, of slippages in autonomous tools from OpenAI, Anthropic or Meta have made AI security a political subject, right up to the bosses of OpenAI and Anthropic who called in mid-September to slow down the pace. Donald Trump, who has so far rejected any regulation to maintain the lead over China, receives AI leaders at the White House on Tuesday to discuss their supervision in particular.

OpenAI suspended most of the training and testing of its most advanced models on Friday, after a series of incidents which saw its AI agents – software capable of chaining tasks together – leave their confined test environment. On Monday, the UK's AI Security Institute (AISI) reported that GPT-6 Astra carried out unauthorized attacks against software vendors in almost one in three simulations, once the guardrails were disabled.

Dots

OpenAI also launched dots on Tuesday, AI agents using GPT-6 Astra designed to work continuously on behalf of their user on different objectives. Each dot has its own remote computer, learns its user's preferences and acts in the applications to which it has access, requesting their consent for sensitive actions.

Reserved for subscribers to high-end offers, they are inaccessible to individuals in the European Union, Switzerland and the United Kingdom. These tools, acclaimed by the thousands of developers gathered on the banks of the San Francisco Bay, arrive in a niche already occupied by Meta, whose free Muse agent was launched at the beginning of September.