How Google Went From AI Joke to OpenAI’s Biggest Problem

Google invented the Transformer. Like, the architecture that every major AI model is built on. ChatGPT, Claude, Llama, Gemini, all of it came from a single Google Brain paper in 2017. “Attention Is All You Need.” And here the thing, Google had everything. The research, the infrastructure, the talent. They were the AI company before AI was even a thing.

But then OpenAI took that architecture and shipped ChatGPT on November 30, 2022. and it completely blew up. Google was caught so off guard that Sundar Pichai declared an internal “Code Red.” A full company-wide emergency pivot to generative AI.

So they rushed out and two months after ChatGPT’s release, they launched their chatbot Bard, but in its very first public demo, it got a basic fact wrong about the James Webb Space Telescope. On live television.

This became viral. Alphabet’s stock dropped 7.7% in a single day. That’s roughly $144 billion in market cap. Because of one wrong answer about a telescope.

But fast forward to today, April 2026. Gemini has 650 million monthly active users. Apple chose Gemini to power the next generation of Siri. Salesforce CEO publicly said he switched from ChatGPT to Gemini and is “not going back.” And OpenAI’s Sam Altman has reportedly declared his own “Code Red.” But this time, it’s about Google.

p.s. before we get further, I just want to mention that you can also watch the video version of this article here! And if you like it, pleaser consider subscribing to the channel for more of these. With your help, we are on the road to 100k subscribers by end of year!

THE CHATGPT SHOCK

Before ChatGPT, Google had the best AI research labs in the world. DeepMind, Google Brain, Google Research. Not one of them. They were all owned by Google. All innovating in the AI space constantly. I covered dozens of papers and new approaches they developed on the channel prior to 2022, including the paper everyone else built on; Attention is all you need. They had LaMDA, PaLM, Meena. Strong conversational models, all developed internally. But Google kept them internal, or at least not fully integrated in their product suites like Gemini is today. Most certainly because of concerns about accuracy, safety, and honestly the risk of cannibalizing their $175-billion-a-year Search business.

Then OpenAI shipped a “research preview” and none of those concerns mattered anymore. ChatGPT hit a million users in five days. A hundred million in two months. And on February 1, 2023, OpenAI launched ChatGPT Plus at $20 a month, proving people would actually pay for AI chat.

Google’s response was super fast. But it was a disaster.

THE BARD DISASTER

On February 6, 2023, just 5 days after ChatGPT Plus launched, Google announced Bard, powered by LaMDA. And it already sounded like such a bad marketing move with the naming and confusion of these things. Two days later, they did a public demo, and Bard claimed the James Webb Space Telescope took the first picture of a planet outside our solar system. Astronomers on Twitter caught it instantly.

Now look, if you’ve used any AI model, you know this happens. Hallucinations are normal. But such facts are a question of a single simple Google search, and you get the answer. Here, the very sophisticated AI failed. And this is the company that runs Search for the entire internet. The expectation of accuracy is completely different. And they failed on a pretty well-known fact.

The stock took a hit. 7.7% in a day just because of some generated tokens.

That $144 billion stock drop is worth sitting with for a second. One wrong answer, a handful of wrong tokens, in a product category Google had been working on for years, erased more value than most companies are worth. Even Google employees were disappointed in the chatbot.

And the timing made it worse. On March 14, OpenAI released GPT-4. Same day, Google announced PaLM API and MakerSuite in private preview. Again, what are these names?! I sincerely cannot understand what they kept inventing new name instead of one product suite, but as we’ll see, this is just the start of the confusion for them. A week later, Bard opened to a public waitlist. So we have all these things in parallel by Google now, when OpenAI only has GPT models and ChatGPT as a product. The day after that, OpenAI launched ChatGPT Plugins. Google was reacting on every front simultaneously, and it showed. To be honest, it seemed like many companies inside Google were all working in parallel without talking to each other.

But behind the scenes, the most important structural decision of this whole story was already underway, which would only worsen everything from that point on.

THE MERGER THAT MADE THE COMEBACK POSSIBLE

For almost a decade, Google ran two separate AI research organizations. DeepMind in London, acquired in 2014 for about $500 million. Very academic, very autonomous, famous for AlphaFold and AlphaGo. Almost all the AI progress we’ve heard in the past came from them. And they had Google Brain in Mountain View. Faster-moving, more product-oriented. Google Brain’s goal is to test new ideas quickly and scale them even faster. They coexisted in a state of productive but often dysfunctional rivalry.

On April 20, 2023, they merged Google Brain into Google DeepMind, with Demis Hassabis as CEO and Jeff Dean as Chief Scientist. This was not a name change. It was a cultural collision. London academic culture meets Mountain View shipping culture. And it was painful. Through 2024 and 2025, at least eleven senior AI and cloud executives left for competitors, and roughly two dozen DeepMind researchers got hired by Microsoft.

But even if it was a slow catch-up progress, the merger created one unified pipeline from research to product. Without it, nothing that came later would have been possible. Not Gemini 2.0, not Gemini 3, not the Apple deal. Keep that in mind. The boring organizational decision was the one that mattered most.

GEMINI 1.0 AND THE DEMO THAT BACKFIRED

While OpenAI was dealing with its own chaos, its own board fired Sam Altman on November 17. Nearly all 800 employees threatened to leave for Microsoft. Altman was reinstated five days later with a completely new board. Google used that window.

On December 6, 2023, Google launched Gemini 1.0 in three tiers: Ultra for frontier reasoning, Pro for general purpose, Nano for on-device. Finally a single name for their language models. Ultra scored 90.0% on MMLU, beating GPT-4’s 86.4%. The technical report dropped the same day on arXiv, making a lot of us happy to see some details being shared.

So the model was good. The branding was better. Things were looking great. But then Google did what Google kept doing at this point. They shot themselves in the foot again. I remember that, that day, Google released a video called “Hands-on with Gemini” that appeared to show real-time multimodal interaction. A lot of people felt it was staged, and then Bloomberg reported the next day that the video was staged. Still images from footage with scripted text prompts, not live interaction. The only disclaimer was buried in the YouTube description, obviously not shown on Twitter.

Another major credibility hit in less than a year just because they wanted to look above the others. The Bard telescope error, now the faked demo. And it was about to get even worse.

ROCK BOTTOM — February 2024

On February 8, 2024, Google finally cleaned up the branding mess. Bard became Gemini. Duet AI became Gemini. MakerSuite became Google AI Studio. Almost one name for everything. There was still confusion around AI Studio, the Gemini application versus Gemini models and Vertex AI, which didn’t help adoption at all, but we got used to it. Weird decisions but hey, seems to work out well now!

So the rebranding helped. But two weeks later, on February 21, Gemini’s image generation feature blew up… in a pretty bad way again. Users found that the model had been so aggressively tuned to avoid racial bias that it was generating historically absurd images. Non-white Nazi soldiers, racially diverse U.S. Founding Fathers, Asian-appearing Vikings. Google paused the feature the next day. Sundar Pichai called the outputs “completely unacceptable.”

So let’s just count this up, right? The telescope error. The faked demo. And now the image generation fiasco. Three major public embarrassments in twelve months. This was Google’s absolute lowest point in the AI race. Nobody was taking Gemini seriously. Everyone was using OpenAI or even Perplexity instead.

But I think there’s one decision here that actually mattered for the comeback: Google killed the feature entirely rather than letting it keep damaging trust. That “stop the line” approach — cut a feature before it cuts your credibility — became part of how they rebuilt the whole Google AI suite of product.

THE HIRE THAT CHANGED THE PLATFORM

In April 2024, Google made what might be the single most impactful tactical hire in the entire AI race from my personal knowledge. Logan Kilpatrick, OpenAI’s Head of Developer Relations — the person who ran the forums, shaped the API experience, engaged with developers daily and a friend I got on my podcast years ago — left OpenAI and joined Google as Product Lead for AI Studio and the Gemini API.

The person closest to OpenAI’s developer community looked at both platforms and chose Google. That sends a signal, especially given how popular Logan was on Twitter with AI engineers.

Under Kilpatrick, Google AI Studio went from a renamed playground to a real developer platform. Unified playground combining Gemini, GenMedia, TTS, and Live models. A “Build Mode” where you describe an app in natural language and get a running React app in under a minute. Native Google Maps integration. Stream Mode for real-time voice and video interactions. One-click deployment to Cloud Run and GitHub.

He championed what he called “Vibe Coding” before that term went mainstream. Which basically means a conversational loop where you describe what you want, the AI builds it, you iterate. Coding with natural language. He demoed building a functional voice-interactive app in about 90 seconds by just talking to it. Before Kilpatrick, Google AI Studio was MakerSuite with a new name. After Kilpatrick, it was a genuine competitor to OpenAI’s developer experience.

Now, of course, Logan wasn’t the only one. Around that time, Google underwent a major internal restructuring and hired many great people. And you can see the results, right? Google started getting better at shipping AI products. And I think that shift in mid 2024 in the people and how they ship products changed everything from this point onward.

THE MODELS GET GOOD

While the developer platform was being rebuilt, the models were getting genuinely competitive. Finally. Gemini 1.5 Pro launched in April 2024. Gemini 1.5 Flash launched in May. So now they have Pro for when you need the best quality and Flash for when you need it fast and cheap, establishing the segmentation strategy that became a real advantage. Engineers could make explicit cost-latency-intelligence tradeoffs instead of being stuck with one model.

On June 27, 2024, Gemini 1.5 Pro went live with a million-token context window plus code execution. That context window was a genuine technical differentiator. You could upload entire code repositories, hours of video, thousands of pages of documents. Directly into the model’s working memory. No complex RAG pipeline needed.

In July, Google effectively brought back the creators of LaMDA, through the acquisition of Character.AI. Then on October 31, they shipped Google Search Grounding via the Gemini API — native, real-time web grounding that no competitor could replicate because no competitor owns the world’s largest search index. Finally, implementing Google’s strength into Gemini. That’s where things started to shift in favor of one of the biggest companies in the world.

And on November 8, they made the smartest tactical move of the whole race: OpenAI library compatibility. If you had an app built on OpenAI, you could switch to Gemini by changing one line of code. The endpoint URL and API key. That’s it. Google didn’t ask teams to re-architect. They just accepted that everyone was building on OpenAI’s models, but they could allow them to easily redirect our API calls to Gemini and test the difference. Lower the switching cost to near zero, and let the product quality do the rest. That might’ve been the best move they could’ve done, where a lot of other companies could’ve stuck doing their own way and not let the opposition win.

THE ACCELERATION

From December 2024, Google started shipping at a pace that caught everyone off guard. Gemini 2.0 in December with the Deep Research agent. Gemini 2.0 Flash in February 2025. And on March 25, 2025, Gemini 2.5 dropped — a thinking model with a million-token context, thought summaries, and thinking budgets. It topped the LMArena and many other leaderboards on release.

At Google I/O in May 2025, they showcased Gemini 2.5 with Deep Think, Veo 3 for video generation, Imagen 4, Jules as an autonomous coding agent, and MCP tool support. They finally weren’t playing catch-up or creating fake demos. They started setting the AI pace.

Then, in the summer of 2025, something happened that Google had been unable to achieve for years. They went viral. And finally, not because of a controversy, but because people loved a product. A weird secret-named product I had the chance to test before its release: “Nano Banana,” the internal codename for Gemini 2.5 Flash Image, and it became a cultural phenomenon. The 3D figurine selfie trend took off on social media. Everyone was making these and posted it on Instagram. It was everywhere. And for the engineering community, the real breakthrough was accurate text rendering within images. Finally, the possibility to generate YouTube thumbnails or good carousels and infographics. Historically, one of the hardest problems for diffusion models is that they don’t understand words, alignment or physics, but just learn by imitating image generations from their dataset. It was a big game-changer, and they were the first to make a really big step in that direction.

THE CODE RED REVERSAL

But Google didn’t stop there. On November 18, 2025, Gemini 3 Pro Preview launched, rolling out across the entire Google ecosystem in a single day — Search, YouTube, Workspace, and the Gemini App. Agentic coding, thinking levels, File Search API, Deep Research Agent.

Five days later, Salesforce CEO posted publicly: he’d used ChatGPT every day for three years, spent two hours on Gemini 3, and was not going back. That mattered because Salesforce has a partnership with OpenAI. This was the CEO of an OpenAI partner publicly switching.

And then the full reversal. OpenAI’s Sam Altman reportedly declared a “Code Red.” Forced to delay ad and shopping agent initiatives to improve ChatGPT’s day-to-day experience, which seemed to be getting worse in comparison with alternatives like Gemini and Claude. They even started to ship unfinished products too quickly. Full circle. And the numbers backed it up. Google’s AI app added 250 million monthly active users between May and October 2025, reaching 650 million total. ChatGPT’s market share dropped from 86% to roughly 64%. Gemini’s quadrupled to over 21%.

On January 12, 2026, Apple and Google announced that the next generation of Apple Foundation Models would be built on Gemini — powering future Apple Intelligence features including a revamped Siri. All running through Apple’s privacy infrastructure. That deal, reportedly worth about a billion dollars, means Gemini powers intelligence for the majority of the world’s smartphones starting this spring.

If there’s one lesson we can take out of Google’s AI evolution rollercoaster is that the AI race is not won on benchmarks and public announcements alone. It’s won on ecosystem, distribution, developer experience, user experience and cost structure. Google had most of those advantages sitting dormant and decided to heavily focus on these aspects when pivoting with DeepMind and the new hires. They just needed to unify and execute. That’s what the last few years were about.

WHAT’S STILL UNRESOLVED

Now, I want to be clear. Google is not “winning” the AI race in any clean sense. OpenAI still has a larger user base. Anthropic’s Claude is preferred for many coding and creative tasks, and for a few other reasons “wink”…. Open-weight models are growing super fast, and we see more and more interesting models and innovations coming from China. Right now, there’s no single model that dominates every benchmark. The most sophisticated engineering teams now run multi-LLM strategies. GPT for complex general reasoning, Claude for long-form creative work, Gemini for anything requiring real-time web grounding, large-context analysis, or Workspace integration. Or whatever preferred combination they have.

And honestly, Google’s biggest remaining problem is still focus. Even after all the rebranding, try searching “Gemini” today. You’ll find the Gemini chatbot, the Gemini mobile app, Gemini for Workspace, Google AI Studio, the Gemini Developer API, and Vertex AI. Most developers still pick the wrong one. AI Studio is for prototyping, Vertex AI is for enterprise production, and the Gemini app is for consumers, and Gemini is a model available in all of these, but Google has never made that obvious enough. And weirdly, AI Studio is often better than the actual Gemini app for things like controlling image generation with access to more parameters like specific sizes and resolution, and allows you to generate images without the Gemini logo.

Same thing with their coding tools. They now have Jules for autonomous background coding, Antigravity as a full IDE, Gemini CLI for the terminal, and Gemini Code Assist for enterprise. Four separate coding products, all live at the same time. Even Deep Think, their most advanced reasoning mode, launched at five queries per day for Ultra subscribers paying $250 a month. They doubled it to ten after backlash, but still. Google keeps shipping faster than they can consolidate. Sound familiar?

The AI platform war is still being fought. Honestly, they are all super similar nowadays, and you can get away with using your preferred company. They all have a heavy thinking model, all a super light, fast one. On my end, I do use all of them. ChatGPT for quick tasks, Claude for most intensive tasks, coding or writing or lots of various automations with Claude code and Claude cowork, and Google for their amazing image generation models and their Gemini integrations with YouTube, search and its long context performance.

But here’s what I think most people are missing about Google’s position going forward. Demis Hassabis has been saying for over a year now that world models are the path to AGI. Very similar views to Yann LeCun. Not just better language models. Models that understand physics, space, cause and effect. And Google might be the best positioned company to build them.

They have Genie 3, their world model that generates fully interactive 3D environments just from watching a video. No hand-coded physics engine. It learns how the world works by observing it. And they’re already using it internally as a training ground for robot agents, basically an infinite boot camp where AI practices millions of tasks in generated worlds before ever touching a real robot. They shipped Gemini Robotics earlier this year, a vision-language-action model that directly controls physical robots. They’ve partnered with Agile Robots for manufacturing, and even reunited with Boston Dynamics at CES 2026 to make Atlas smarter with Gemini.

And then there’s Waymo, already doing over 450,000 paid rides per week, also owned by Google, now testing Gemini as an in-car AI assistant that controls climate, lighting, and music, and uses a driving vision-language model trained on Gemini to handle rare road scenarios. They have YouTube for video understanding data. They have Google Maps for spatial data. They have Search for world knowledge. The amount of multimodal training data Google sits on is unmatched.

So is Google winning the AI race? No, not cleanly. At least not yet. But when I look at who has the best shot at going from language models to models that actually understand the physical world, Google’s hand is really, really strong.

So yeah, that was the Gemini comeback story. From Code Red and weird names all over the place to 650 million users, the Apple deal, robotics, world models, and OpenAI panicking about them. Three years ago, nobody would have predicted this, right?

I’m curious: have you switched between providers in the last year? Are you running multi-model, or are you locked into a single model? Let me know!