Skip to content
Artificial Intelligence

Anthropic Could Be the Real Winner in OpenAI’s Latest Model Release

The AI influencers aren't happy with OpenAI's new update.
By

Reading time 3 minutes

With Anthropic’s IPO expected tomorrow, OpenAI really needed to wow users with its latest model release. Judging from the early reviews, however, GPT-6.2 had exactly the opposite effect. 

Released Tuesday, the new model is being sold by OpenAI as an expert digital assistant for real-world projects in fields like finance, law, and accounting. It comes in two varieties: GPT-6.2 Astra—which the company describes as its “frontier model for your hardest projects”—is supposed to excel at drafting documents, slideshows, and other materials that don’t smack of being AI-generated and that closely adhere to templates and design specifications provided by the user. (Echoing a similar announcement from Anthropic earlier this month, OpenAI claimed in its announcement that GPT-6.2 Astra is its first publicly released model to cross its “High threshold for AI self-improvement”—i.e., recursive self-improvement—a designation which it said was roughly equivalent to providing human AI researchers with a “highly capable” assistant.) The faster, more efficient GPT-6.2 Sol is a specialist in handling everyday tasks using fewer tokens than its predecessor, GPT-6.1 Sol. OpenAI wrote in its announcement that many of its employees use GPT-6.2 Sol “as their default model for both technical and non-technical tasks.”

Prior to yesterday’s announcement, OpenAI let several well-known tech commentators take the new model for a spin. Many of those early testers published their reviews yesterday. And for the most part, they were eviscerating—both of the model itself and of the decision-makers within OpenAI who signed off on its public release.

Entrepreneur and tech investor Matt Shumer initially wrote in an X post that his experience with GPT-6.2 Astra and Sol had been so discouraging that he’d decided not to publish a formal, in-depth review. “They were essentially unusable,” he summed up curtly. Spurred on by his followers, he eventually capitulated and posted a 14.5-minute scathing review on YouTube. Shumer’s early tests with the model left him with the impression that it was a major step back in terms of capability and user experience. The major frustration he kept returning to was that it would incessantly ask for permission on innocuous steps, such as whether or not it should download a file—steps that most models currently on the market would just go ahead and handle.

“This model has no concept of what autonomy means,” Shumer said in his review. “It’s dramatically worse than the previous OpenAI models, but it is leagues behind current Anthropic models.” He went on to concede that with a little meticulous post-training, its successor could be much stronger. “But right now, it feels like OpenAI handed Anthropic a gift right before their IPO.”

McKay Wrigley, a software developer and the CEO of the AI startup Takeoff, echoed that sentiment in his own brief review: “astra 6.2 is a pretty good ad for anthropic’s ipo,” he wrote on X. Ethan Mollick, an associate professor at the Wharton School at the University of Pennsylvania, praised GPT-6.2 Astra’s performance after he prompted it to generate a simulation of a space elevator using 3D animating platform Blender and the game engine Unity; but he too said it “required more guidance and pushing to get work done” compared to other models and “was often frustrating to use.”

So negative was the first wave of reviews that OpenAI product lead Thibault Sottiaux, better known as Tibo, posted a reply in defense of the model. AI “influencers” were given access to GPT-6.2 “a while ago,” he wrote, but within the three or four days before its release, his team had made a number of changes. Sottiaux suggested that the early reviews might not accurately reflect the models’ capabilities at launch.

If it seems a bit strange that OpenAI would give early testers—each of them with hundreds of thousands of online followers—the keys to its new model and then make “a lot of changes” (Tibo’s words) at the eleventh hour, and apparently without giving them a heads up, there are some who’d agree. 

Entrepreneur and AI commentator Dan Shipper, who posted his own less-than-glowing review of GPT-6.2—describing it as “a disappointing release compared to [Anthropic’s] Fable 5.2” and specifically calling out shortcomings in writing, coding, and design—responded to Tibo’s post claiming that his company had in fact continued testing the model within the timeframe the new changes had supposedly been made. “I don’t see any comms from the @openai team that there were new changes to test over the last 3-4 days,” he wrote. “We absolutely would’ve started from scratch if there were!”

OpenAI did not immediately respond to Gizmodo’s request for comment.

Explore more on these topics

Share this story

Sign up for our newsletters

Subscribe and interact with our community, get up to date with our customised Newsletters and much more.