Google's Gemini Spark: Promising AI, but at What Cost?

Gemini Spark: A Promising but Costly AI Agent

Google recently introduced Gemini Spark, an artificial intelligence agent designed to handle complex tasks in the background. While its capabilities seem impressive, its high cost and privacy trade-offs raise questions.

Last week, a test of Spark was carried out, with Google presenting the assistant as capable of managing multiple tasks while remaining under user control. Google emphasizes that Spark requires manual activation and checks with the user before making significant decisions, a precaution in response to growing skepticism towards AI.

Not knowing where to start, the tester decided to test Spark by performing tasks similar to those Google demonstrated during the I/O event. Would it work as well in a home office as it did on the big stage?

Testing Gemini Spark

During I/O, Google's vice president, Josh Woodward, showcased several examples. The first involved asking Spark to draft an email to a Google team, compile all information on the launches of Gemini Live and the "successes from last week," and use a special AI skill to make the email sound like him. The tester therefore decided to push Spark a bit further.

Gemini was asked to draft an email to the tester's wife compiling their average monthly grocery expenses for 2026. This test would allow several things to be checked: could Spark identify the tester's wife (without her name being given), could it locate their budget sheet in Drive (which does not contain "budget" in the file name), and could it actually draft an email in Gmail?

When Spark's output was received shortly after, the reaction was: "Wow, that's crazy." Spark found the wife's email address, extracted the correct information from their 2026 budget sheet, retrieved the monthly grocery totals, including incomplete data from May (which wasn't finished when the test was conducted), calculated the average of the totals, and put everything into a draft email in the tester's Gmail. The email text mentioned the wife by her first name, even though her email address did not contain her first name. It even included a polite closing used just between them.

In his next example, Woodward asked for help planning a neighborhood party. As the tester was not planning a neighborhood party, he asked Spark for assistance by posing the same questions Woodward had asked. It didn't go well. Spark created a table of friends and family as a "very realistic reference for who brings what," drafted an email in the tester's Gmail mentioning a shared sign-up sheet that doesn't exist, and created an unappealing presentation with slides detailing information about city permits.

To push Spark further, it was asked to create that missing sign-up sheet and add a link to the already drafted email. Although Spark took a few minutes to understand, this task worked; it created a spreadsheet and returned to the draft email text to insert the link.

Woodward's final demonstration was undoubtedly the most impressive. He asked Spark to do several things: color his meetings with CEO Sundar Pichai bright pink on his calendar, draft a note to a new neighbor inviting them to his neighborhood party, and create a document to help with tasks for his children at the end of the school year. For his version, the tester asked it to create a calendar event each month before his wife's birthday and make it bright pink, to draft an email to his family sending them the first episode of the latest season of Taskmaster, and to create a document with the key things he and his wife need to know to prepare their toddler for kindergarten.

The request was started at 3:35 PM PT on a Friday. During I/O, Woodward put on a little show by setting his phone down and promising to check the results later in the keynote, which he did. But after a small issue was resolved (Spark wanted access to the tester's contacts, which was denied), the task was completed about four minutes later.

Once again, the tester was somewhat impressed by the results, although they were imperfect:

• The tester's Google Calendar now has events from 9 AM to 10 AM on the correct day of each month leading up to his wife's birthday. The reminders are in what Google calls "flamingo," which isn't exactly "bright pink," but close enough.

• Spark retrieved the emails of the tester's immediate family and put them into a draft email. (Strangely, it didn't include his wife's.) The email text correctly mentioned the name of the first episode of the latest season of Taskmaster, but linked to a trailer instead of the actual episode. The email also contained the term "loool," which the tester uses in casual written conversations.

• Spark created a Google Doc in the tester's Drive with a checklist for kindergarten preparation. However, it is only accessible to the tester; Spark was asked whether it could grant access to the tester's wife, but it replied that it currently could not.

Conclusion

Spark could be a powerful tool. But there are a few caveats to mention. Like all AI tools, its results must always be verified to ensure their accuracy, which could carry higher stakes when it comes to personal information to be shared with people one knows. Although Google presents Spark as something that can operate autonomously, the tester found himself constantly monitoring it or checking the notifications it sent to his phone. What's the point of an assistant if every move has to be micromanaged instead of trusting it? And why should something that feels so uncertain drain energy from a resource-hungry data center for relatively insignificant tasks?

Currently, Spark is only available to subscribers of Google's AI Ultra plan, which starts at $99.99 per month, and only for users in the United States and in English. Google provided free access to test Spark, and the tester does not think it is good enough to justify these expensive plans. Especially when all the tasks asked of Spark could be accomplished by the tester himself (they would just take longer).

Spark also works better if the user is already well integrated into the Google ecosystem and has Personal Intelligence enabled. The tester has had a Google account for about two decades, so Spark has a lot of data it can use to inform its responses. But while Google promises that Gemini "does not directly train" on the Gmail inbox with Personal Intelligence enabled, Google still has to be trusted to be a good steward of the data. For now, the tester is not sure it is worth the cost or the risk.