Is Google Losing the AI Race? What's Really Going On With Gemini
Demis Hassabis's exit and shaky Gemini releases fueled claims Google is falling behind. Here's what's actually happening at DeepMind.
Is Google actually losing the AI race?
Not clearly, but the last few months have made a strong circumstantial case. Google shipped Gemini 3 Pro in November and briefly had, by most accounts, the best model in the industry, reportedly forcing OpenAI into an internal “code red.” Then the follow-up models underwhelmed, key executives announced departures, and Google’s newest text and coding models stopped showing up near the top of major leaderboards. None of that proves Google is out of contention. It does show a company that lost momentum right after it had seized it.
What happened with Demis Hassabis and Google DeepMind leadership?
The trigger for the current round of “Google is falling behind” discourse was leadership news, not a benchmark. Demis Hassabis is stepping back from his role as CEO of Google DeepMind to become the unit’s chairman and chief scientist. Around the same time, Jeff Dean and another senior Google AI executive announced they’re leaving to start their own AI company, one that Google is reportedly going to invest in. Noam Shazeer and John Jumper, both widely regarded as among the strongest researchers in the field, have also been named as recent departures.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Taken together, that’s a lot of senior talent moving away from the day-to-day model-building operation in a short window. It’s easy to read that as a company in retreat. But there’s a more specific explanation for at least one of these moves: Hassabis has talked for years about wanting to use AI to accelerate scientific discovery, particularly in medicine. Stepping into a chairman and chief scientist role while leading Isomorphic Labs, Google’s AI drug discovery spinoff, lets him pursue that directly instead of running a lab that has to ship competitive consumer products on a quarterly cycle. That’s a mission shift, not necessarily a vote of no confidence in Gemini.
How far behind is Gemini on the benchmarks?
The picture is uneven depending on modality.
On image generation, Google’s Nano Banana family remains fast and broadly useful, but it no longer tops the text-to-image leaderboards. It has been passed by models from Meta, Microsoft, xAI (Grok), and others, landing around seventh on at least one major arena ranking rather than in the top few spots. A few months ago, a prediction that Meta would out-rank Google on image generation would have sounded unlikely. Now it’s the case.
Video is a brighter spot. Google’s video model has been praised for speed and for handling certain tasks competitors can’t replicate as well, and it’s plausible it still ranks near the top among Western labs, even if a newer Chinese model may have since taken the overall top spot on some boards.
Text and reasoning models are where the concern is sharpest. At the point this debate flared up, Google’s most current model in circulation was Gemini 3 Flash, and it didn’t rank near the top of leaderboards for math, software engineering, coding, or cybersecurity, categories where labs like OpenAI and Anthropic have been pulling ahead. Industry chatter (notably from the analyst outlet SemiAnalysis) suggested Google might skip a “Gemini 3.5 Pro” release entirely and jump to Gemini 4, with speculation that an interim model performed only around the level of Anthropic’s Opus 4.5, behind competing models like GLM 5.2 and well behind GPT 5.6 territory. SemiAnalysis went further, arguing Google may lack what it called the “religious conviction” needed to push toward recursive self-improvement (RSI), the kind of aggressive, all-in bet on scaling that rivals are perceived to be making.
That’s a pointed claim from a research outlet that tracks compute, model releases, and lab strategy closely and has a track record of predictions holding up. It’s still one outlet’s read, not a settled fact.
Is there a case that Google isn’t actually in trouble?
Yes, and it rests on three things: infrastructure, timing, and a pattern seen across the industry.
Google engineer and Gemini team member Logan Kilpatrick pushed back publicly on the SemiAnalysis framing, calling it superficial and saying the Gemini team is “cooking” with talented people doing serious work. That’s an interested party defending his own team, so it’s worth discounting somewhat, but it’s not nothing either.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
More concretely, Google has responded organizationally. Sundar Pichai has signaled leadership changes at DeepMind. Sergey Brin, Google’s co-founder and one of the wealthiest people alive, has reportedly come out of semi-retirement to take direct command of Gemini development, working hands-on rather than just overseeing from a distance. That mirrors moves other AI leaders have made when they felt their company was losing an edge: Mark Zuckerberg went into what’s been called “founder mode” and reorganized Meta around a superintelligence push, and Elon Musk has run xAI with similarly hands-on intensity. Founders parachuting back into the trenches is usually a sign a company is trying to fix something quickly, not a sign of giving up.
There’s also a newer data point cutting against the doom narrative: Gemini 3.7 Flash reportedly beat several frontier-class models, including ones from Anthropic and OpenAI, on an agent benchmark from Artificial Analysis. It’s a smaller, cheaper “Flash” model rather than Google’s flagship, which makes the result more notable, not less. Reports also point to it performing well on ARC-AGI-2 relative to its cost, and to strong results on video analysis tasks specifically.
Has this kind of slump happened to other AI labs before?
Repeatedly. xAI has had stretches where it looked outmatched before releasing a model that reset expectations. OpenAI has had quiet periods followed by a release that reasserted its position. Anthropic and Meta have both had moments where they looked like they’d lost the thread, only to come back with something competitive. The AI industry currently moves in cycles measured in months, not years: a lab ships a genuinely strong model, competitors adjust and catch up or pass it, and then the original lab ships again. Gemini 3 Pro’s brief run at the top in November, followed by a rockier stretch, fits that same rhythm rather than breaking it.
Google also has structural advantages that don’t show up on a leaderboard: distribution through Search, Android, Chrome, and Workspace, plus its own TPU hardware, which reduces reliance on external chip supply. Those don’t guarantee the best model, but they mean Google can put a competitive model in front of more people, more cheaply, than almost anyone else the moment it has one.
Frequently Asked Questions
Is Google behind OpenAI and Anthropic right now?
On several text, coding, and reasoning benchmarks, Google’s most recent flagship-tier releases have not ranked at the top, and image generation has also slipped from its earlier lead. Video generation and some agent benchmarks (via Gemini 3.7 Flash) are areas where Google still looks competitive or ahead.
Why is Demis Hassabis stepping back from Google DeepMind?
He’s moving to chairman and chief scientist of DeepMind while focusing on Isomorphic Labs, Google’s AI drug discovery spinoff, which aligns with his long-stated interest in using AI to accelerate medical research rather than running day-to-day frontier model shipping.
Did Google really have the best model in the world recently?
According to reporting cited by analysts, Gemini 3 Pro was considered arguably the best model in the world around November, reportedly prompting an internal “code red” response at OpenAI, before Google’s subsequent releases underperformed relative to that peak.
Why is Sergey Brin getting involved with Gemini directly?
Brin has reportedly come out of retirement to take direct, hands-on command of Gemini development, a move comparable to Mark Zuckerberg’s “founder mode” push at Meta or Elon Musk’s direct involvement at xAI when those companies wanted to accelerate their AI efforts.
Could Google catch back up to the frontier?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The transcript’s analysis, and the broader pattern in AI, suggests slumps like this have hit xAI, OpenAI, Anthropic, and Meta before, each followed by a comeback release. Combined with Google’s distribution advantages and renewed leadership focus, a turnaround is plausible, though not guaranteed.