A year ago, the era of generative AI chatbots was new, and sitting around with your mouth agape was a perfectly understandable reaction. OpenAI’s ChatGPT was amazing! In some ways, Bing Chat—which drizzled Microsoft special sauce on the same large language model as ChatGPT—was even more amazing! As for Google’s Bard . . . well, it was slightly less amazing than ChatGPT and Bing. But given Google’s formidable AI chops, there was every reason to think it might catch up quick.

As the sheer novelty of these chatbots has faded, a more mundane matter has emerged: Are they worth paying for? That question moved back to the front of my brain last week when Google introduced a new chatbot aimed squarely at the folks who may currently be subscribing to OpenAI’s ChatGPT Plus.

Google’s paid bot, Gemini Advanced, is part of a new tier of its Google One service and arrives amid a rebranding of the company’s existing AI offerings. (The free bot formerly known as Bard is now plain old Gemini; Duet AI, a collection of business-focused tools, is in the process of being renamed Gemini for Workspace.) For $20 a month, Google One AI Premium offers Gemini Advanced, 2 TB of Google Drive cloud storage, extra features for Google Photos and other apps, and security aids such as a VPN. Subscribers will also be able to use Gemini Advanced inside Gmail, Google Docs, and other apps; that integration is “coming soon,” Google promises.

All this costs $20 a month, matching what you’d pay for ChatGPT Plus alone. Google is even offering an uncommonly generous two-month free trial.

Under the surface, Gemini Advanced leverages Gemini Ultra 1.0, Google’s most powerful LLM. Wharton professor Ethan Mollick says it’s “roughly equivalent [to OpenAI’s GPT-4], though it has its own strengths and weaknesses.” He knows as much as anyone about how major AI products compare. But my first few days with Google’s new bot didn’t leave me tempted to reallocate the $20 I’m currently paying for GPT Plus. (Full disclosure: Fast Company is nice enough to reimburse me, but I’m not about to expense two $20 chatbots per month.)

In multiple respects, Gemini Advanced struck me as a buggy beta, not a finished product. Oddly enough, I usually couldn’t use it at all until I’d completed a CAPTCHA to prove I wasn’t a bot. Even after I followed the instructions to shut off the ability of human Google employees to read my prompts as a quality-control aid, I continued to get alerts informing me they had access to everything I entered. There’s a feature for verifying responses via Google search, which sounds invaluable but never worked for me.

ChatGPT Plus now begins to answer my prompts almost instantly; Gemini Advanced chugs away for 5 to 15 seconds before saying anything. Some of its replies to my random queries, on topics ranging from evicting raccoons to the difference between interpreted and compiled computer languages, were cogent and helpful. Unlike ChatGPT, it weaves images into its responses—usually pertinent ones, though it can be flummoxed by complications such as multiple people sharing one name—and can handle visual requests such as “Show me the Puerto Rican flag.”

Compared to ChatGPT’s largely anodyne personality, Gemini Advanced leans toward a synthetic chumminess that can get awfully wordy: Even when it was explaining why it couldn’t fulfill my requests, it went on and on, alternately apologetic and defensive. But the bot’s chatty tone didn’t mask the constant reminders that it was stringing words together without comprehending what it was saying.

For instance, two days before the Super Bowl, Gemini Advanced said it was too early to know where I could stream the game. It told me that John Quincy Adams and his wife, Louisa, were the first U.S. president and first lady to be photographed together, and then, moments later, declared that no verified photos of Louisa existed at all. Asked to recommend 1950s British comic novels, it raved about one published in 1889, allowing that it “technically” didn’t qualify. It claimed to be able to send emails and dictate text aloud—neither of which it can actually do—and sometimes insisted it couldn’t generate images, even though it can.

All of this is on top of Gemini Advanced’s more typical AI hallucinations, which seemed greater in number to me than with ChatGPT Plus in its current incarnation. Google pitches its bot as a personal tutor that can test your knowledge with quizzes, but the ones it generated for me were pockmarked with errors a half-competent human teacher wouldn’t make. And the free version of Gemini was even more mistake-prone, to such an extent that I wouldn’t assume anything it says is accurate.

Now, Google many not see treating Gemini like an omniscient encyclopedia as the principal use-case scenario. Like most AI purveyors, it warns you about its bot’s imperfections (“Gemini may display inaccurate info, including about people, so double-check its responses”). I also noticed that the usage suggestions Google listed at the start of new sessions involved activities such as coding, generating images, game playing, and fiction writing—ones for which Gemini’s slippery grasp on facts might be less of a problem, or could even be a virtue.

Regardless of the precautions I took, the bottom line was that I felt like Gemini Advanced was a tool that could blow up in my hands unless I treated it gingerly. ChatGPT Plus can be that way at times, too: I cheerfully admit I get more out of it as an experimental plaything than a workaday productivity app. Still, in close to a year of paying for OpenAI’s bot, I’ve seen it grow more dependable and less weird, in ways that leave it well ahead of Gemini Advanced.

Over the long haul, however, Google may have an advantage in its control of a sizable chunk of the workplace software market. By the time Gemini Advanced’s two-month free trials start running out, it’s possible Google will have launched its planned integrations into Gmail, Google Docs, and other apps (there are already some basic ones in the Gemini app for Android and Google app for iPhone). The utility of applying AI to your own email, files, and other data sources could be worth 20 bucks a month all by itself.

Though not necessarily. On Tuesday, The Wall Street Journal’s Tom Dotan reported that some early adopters of Microsoft’s Copilot chatbot for Microsoft 365 (née Office) are finding it too hallucinatory and generally unreliable to justify the monthly price tag of $30 per user. It turns out that being amazing is not the same thing as being useful—and the latter bar may be tougher for these bots to meet.