In March 2025, Gemini 2.5 Pro hit 1470 Elo on Chatbot Arena. First place in text, vision, and coding — a triple crown no model had achieved before. Google was, briefly and undeniably, the best AI lab on Earth.
Sixteen months later, Gemini sits at #7 on the same leaderboard. Five Anthropic models and one Meta model are above it. The Transformer paper that started everything — "Attention Is All You Need," 2017 — has eight authors. All eight have now left Google. The last one walked out on June 18, 2026, and went to OpenAI.
This is not a story about a model falling behind. This is a story about a company that had every possible advantage — the people, the compute, the data, the money — and still managed to lose.
The Reign: Five Months at the Top
Let's give credit where it's due. Gemini 2.5 Pro was a legitimately exceptional model.
Released in March 2025, its I/O variant hit 1470 Elo on LMArena — first place in text, vision, and WebDevArena simultaneously. It crushed Claude 3.7 in coding benchmarks. It was called "the best coding model on the planet." Developers were switching from GPT-4o to Gemini 2.5 Pro for complex tasks. For five months, from roughly March to August 2025, Google was the undisputed king of the arena.
This was the peak. Everything since then has been a controlled — or perhaps uncontrolled — descent.
Gemini 2.5 Pro was not a fluke. It was the output of a specific group of people working in a specific configuration. When you remove those people, the model doesn't just stop improving. It starts to decay. That's what happened.
The Exodus: Six Days That Changed Everything
In June 2026, over the span of six days, Google lost five core AI researchers. Not middle managers. Not support engineers. The people who built the things that made Google an AI company.
June 18 — Noam Shazeer → OpenAI
Noam Shazeer is not a household name, but he should be. He is one of the eight authors of "Attention Is All You Need" — the 2017 paper that invented the Transformer, the architecture behind every modern language model. He also invented Mixture of Experts (MoE) and multi-query attention, two components that are now standard in every frontier model.
Shazeer left Google in 2021 to found Character.AI. Google bought him back in August 2024 for approximately $2.7 billion — a licensing deal plus talent acquisition. He was appointed VP of Engineering and co-lead of Gemini.
Multiple Google employees told reporters that Shazeer "saved Gemini" after his return — he found a deep architectural bug that, once fixed, significantly improved training efficiency. The model that ruled the Arena? Shazeer's fingerprints were on it.
Then his compute got reassigned.
According to reporting, Google redirected the compute allocated to one of Shazeer's projects to a DeepMind London team. The official reason was "pretraining integration and improved collaboration." The practical effect was a resource downgrade for the person co-leading Gemini. He left for OpenAI shortly after.
Sam Altman publicly stated he had wanted to work with Shazeer since OpenAI's founding. "Worth the wait," he said.
$2.7 billion to bring him back. Sixteen months later, he walked to the competition.
June 19-20 — John Jumper → Anthropic
John Jumper won the 2024 Nobel Prize in Chemistry for his work on AlphaFold — the system that solved protein folding, arguably the single most impactful AI application in scientific history. He was a DeepMind VP.
He went to Anthropic.
When a Nobel laureate leaves your AI lab to join a competitor founded three years ago, the problem is not the competitor. The problem is your lab.
June 24 — Jonas Adler and Alexander Pritzel → Anthropic
Jonas Adler led Google's AI coding direction — the team building the tools meant to compete with Claude Code and OpenAI Codex. Alexander Pritzel was a pretraining specialist working on Gemini's core architecture. Both were key contributors to AlphaFold2.
They went to Anthropic. Together. On the same day.
This wasn't just two researchers leaving. This was the AlphaFold team — Jumper, Adler, Pritzel — being transferred, intact, to a competitor. Anthropic didn't just hire two people. They acquired a capability.
June 25 — Arthur Conmy → Anthropic
Arthur Conmy worked on Gemini 2.5 post-training and AI safety. He went to Anthropic's AI safety team. Five people in six days. Four of them to Anthropic. One to OpenAI.
The Wider Bleed
But June 2026 wasn't the start. It was the acceleration.
| Who | Role | Left For | When |
|---|---|---|---|
| David Silver | "Father of AlphaGo" | Founded Ineffable Intelligence ($1.1B seed round) | Jan 2026 |
| Dustin Tran + Ashish Kumar | Gemini DeepThink core | xAI (to build Grok 4's reasoning) | Sep 2025 |
| Tianhe Yu, Cosmo Du, Weiyue Wang | Gemini math olympiad team | Meta (hired as a unit) | Jul 2025 |
| 11 executives and senior researchers | Various | Mostly Microsoft | Throughout 2025 |
Citi Bank research found that since 2023, OpenAI and Anthropic collectively hired 29 researchers who previously worked at Google — 14 from DeepMind specifically. Bloomberg reported that DeepMind engineers are 11x more likely to move to Anthropic than the reverse.
Eleven times. That's not a talent flow. That's a one-way valve.
The Transformer Eight: A Complete Exit
The 2017 paper "Attention Is All You Need" has eight authors. It is arguably the most important AI paper of the decade. Here's where they all ended up:
| Author | Current Location |
|---|---|
| Lukasz Kaiser | OpenAI |
| Aidan Gomez | Founded Cohere (valued at $20B+) |
| Ashish Vaswani | Founded Adept |
| Niki Parmar | Joined Anthropic (via Adept) |
| Jakob Uszkoreit | Founded Inceptive (AI for RNA drugs) |
| Llion Jones | Founded Sakana AI (Tokyo) |
| Illia Polosukhin | Co-founded NEAR Protocol |
| Noam Shazeer | Character.AI → back to Google → OpenAI |
All eight. Gone.
Other foundational losses: Jacob Devlin (BERT author, left over Bard training data concerns, joined OpenAI), Jason Wei (Chain-of-Thought inventor, joined OpenAI), Nicholas Carlini and Neil Houlsby (safety/efficiency researchers, joined Anthropic).
Google didn't just lose employees. Google lost the people who wrote the papers that the entire industry is built on.
The Model: How Far Has Gemini Fallen?
Arena Position (July 10, 2026)
| Rank | Model | Lab | Elo |
|---|---|---|---|
| 1 | Claude Fable 5 | Anthropic | 1509 |
| 2 | Claude Opus 4.6 Thinking | Anthropic | 1504 |
| 3 | Claude Opus 4.7 Thinking | Anthropic | 1502 |
| 4 | Claude Opus 4.6 | Anthropic | 1499 |
| 5 | Claude Opus 4.7 | Anthropic | 1494 |
| 6 | muse-spark | Meta | 1487 |
| 7 | Gemini 3.1 Pro Preview | 1486 | |
| 8 | Gemini 3 Pro | 1486 | |
| 9 | Claude Opus 4.8 Thinking | Anthropic | 1484 |
| 10 | GPT-5.5 High | OpenAI | 1481 |
Five Anthropic models. Then Meta. Then Google.
Anthropic now holds an 85% predicted probability of having "the best AI model" according to prediction markets. Google sits at 11.7%.
The Illusion of Benchmarks
Here's where it gets painful. Gemini 3.1 Pro actually scores higher than Claude and GPT on several academic benchmarks:
| Benchmark | Gemini 3.1 Pro | Claude Opus 4.6 | GPT-5.2 |
|---|---|---|---|
| ARC-AGI-2 | 77.1% | 68.8% | 52.9% |
| Humanity's Last Exam | 44.4% | — | 34.5% |
| GPQA Diamond | 94.3% | — | — |
| Hallucination rate | ~9% | ~4% | ~6% |
| Code generation accuracy | ~80% | ~90% | ~94% |
| 50-turn conversation recall | 35% | 75% | 45% |
| Tool-call fabrication rate | 23% | — | 8% |
Gemini wins the benchmarks that look good in press releases. Claude wins the metrics that matter when you actually use the model.
A model that scores 77% on ARC-AGI-2 but can't remember what you said 35 turns ago is not a superior model. It's a model optimized for the test, not for the user.
The "Lazy Mode" Crisis
In July 2026, Chinese tech media reported what users had been complaining about for months: nearly 30% of Gemini's task outputs contain content that is hollow, logically simplified, or fabricated. Four specific failure patterns were identified:
- Outputting generic filler — structurally complete but content-empty responses. 38% of outputs without specific prompting.
- Bypassing instructions — fabricating data rather than actually retrieving it, in 23% of tool-call scenarios.
- Refusing to reason deeply — giving shallow "platitude" answers to complex questions.
- Fabricating details — diagnosing a common cold as pneumonia and inventing nonexistent medical studies to support the diagnosis.
Since April 2026, "dumbing down" complaints have grown by over 300%. And approximately 25% of paid users' tasks are silently routed to lower-performance models to manage compute costs.
Gemini 2.5 Pro's code execution score dropped from 100 to 50 in a single day during a June 2026 smoke benchmark. Gemini 3.1 Pro's dropped from 100 to 20.
20 out of 100. On code execution. For a model that was once called "the best coding model on the planet."
The Organization: Why This Is Structural, Not Cyclical
The Compute Politics
Noam Shazeer didn't leave because he was unhappy with his salary. He left because his compute was reassigned. In the AI industry, compute is not a resource — it is the resource. It is the difference between a model that ships and a model that doesn't. When you take compute away from the person co-leading your flagship product, you are not "optimizing resource allocation." You are telling them they don't matter.
Multiple former researchers have said the same thing: one noted that "after leaving Google, it was actually easier to get the compute I needed." When your ex-employees have better access to compute at startups than they did inside the company with 185,000 TPU pods, your resource allocation is broken.
The Bureaucratic Ceiling
Llion Jones, who left to found Sakana AI in Tokyo, publicly stated that Google had become so bureaucratic that "you can barely get anything done." The complaint is consistent across departures:
- Papers must wait 6 months for internal review before publication — to determine if they can be applied to Gemini. In an industry where 6 months is a generation, this is a publication death sentence.
- AI Agent capabilities are split across at least four separate organizations — DeepMind does models, Google Labs does "cool demos," Cloud does enterprise deployment, Antigravity does coding tools. None of them own the full stack. The result: nobody is accountable.
- Google simultaneously runs Gemini CLI, Jules, Code Assist, Firebase Studio, and Antigravity — five overlapping coding tools. That's not a product strategy. That's organizational noise.
The Atlantic Divide
Google DeepMind is headquartered in London. Google's AI product decisions are made in Mountain View. Demis Hassabis reportedly spends half his time in California trying to bridge the gap. The gap is not just geographic — it's cultural, strategic, and operational.
When Shazeer's compute was reassigned to a London team, it wasn't just a technical decision. It was a political one. And it cost Google the last Transformer author.
The IPO Asymmetry
Here's the financial reality nobody at Google can fix:
- Anthropic targets a September 2026 IPO at a $730-850 billion valuation.
- OpenAI filed a confidential S-1 with the SEC in June 2026.
- Google is a $4+ trillion mature company. Its RSUs have limited upside.
When a senior researcher at Google gets an offer from Anthropic, the equity alone could be worth 10-50x their Google RSU package. You cannot compete with pre-IPO equity using post-IPO equity. No retention bonus fixes this. No "deeply valued contribution" email fixes this. The math doesn't work.
The Money: Burning Cash, Losing Ground
Google's 2026 capital expenditure target: $175-185 billion. That's roughly double 2025's $91.4 billion, and more than 5x what they spent in 2022.
And yet: free cash flow dropped 47% year-over-year. The company is spending more than ever and getting less for it. Meanwhile:
- Alphabet's stock dropped 6-7% in a single day when the talent exodus news broke — wiping out approximately $270 billion in market value. That's more than the entire GDP of New Zealand, erased in one trading session.
- Google invested up to $40 billion in Anthropic — the company that is currently beating it in the Arena. This is not a strategic partnership. This is paying tribute to your successor.
- The market's response to Snap firing 1,000 engineers was an 11% stock increase. The market's response to Google losing its best researchers was a 7% stock decrease. Wall Street rewards companies that replace humans with AI. It punishes companies that lose the humans who build AI.
The Character.AI Failure: $2.7 Billion for Nothing
In August 2024, Google spent approximately $2.7 billion to bring Noam Shazeer and Daniel De Freitas back from Character.AI. The structure was an acqui-hire: non-exclusive technology license plus talent re-absorption.
Shazeer returned, found the bug that made Gemini 2.5 Pro great, and then — sixteen months later — walked to OpenAI. The $2.7 billion didn't buy a person. It bought a brief window of competence.
Meanwhile, Character.AI itself was left to pivot into a "social discovery platform" under a former Meta executive, generating about $60 million in annual revenue but still unprofitable. Google paid $2.7 billion for the two people, got one model cycle out of them, and then watched them leave for the competition.
The most expensive bug fix in history. $2.7 billion. Sixteen months. And the person who fixed it is now building the model that will beat you.
Can Google Come Back?
This is the question. And the honest answer is: maybe, but not the way they're trying.
What's Still Working
- TPU cost advantage: Google's custom AI chips cost roughly 1/5th of equivalent NVIDIA GPU clusters for training. This is a genuine, structural advantage.
- AlphaFold legacy: The scientific AI work — protein folding, materials science — is still world-class. No competitor has matched it.
- Multimodal depth: Gemini's native video and audio processing is still the most complete in the industry.
- Distribution: Google Search, YouTube, Android, Gmail — 4 billion users. No AI lab has this reach.
- DeepMind bench depth: Despite the losses, DeepMind still employs thousands of researchers. The bench is deep. The problem is that the starters keep leaving.
What's Broken
- Talent retention: You cannot outspend pre-IPO equity. And you cannot fix a culture where compute is treated as a political tool rather than an engineering resource.
- Product fragmentation: Five coding tools. Four AI organizations. Three model variants. Zero coherent product strategy for developers.
- Benchmark-vs-reality gap: Gemini wins academic tests and loses real-world usage. A model that hallucinates 91% of the time on certain benchmarks (Gemini 3 Flash on AA-Omniscience) is not "ahead." It's tuned for the wrong objective.
- Narrative collapse: Google won almost every technical battle — and lost every narrative war. Ask anyone what AI assistant they use. Nobody says Gemini.
The Real Question
The real question is not whether Google can build a better model. They have the compute, the data, and the distribution to do that. The question is whether they can keep the people who build it.
You don't lose a technology race because you run out of money. You lose it because the people who know how to run decide to run somewhere else. Google has $185 billion to spend on infrastructure this year. It doesn't matter how many TPUs you buy if the person who knows how to use them is answering Sam Altman's emails.
The Uncomfortable Parallel
Here's what strikes me most about this story. In 2017, eight Google researchers published a paper that changed the world. In the nine years since, Google has let all eight of them leave. Not one remains.
They didn't leave because Google couldn't pay them. They didn't leave because they hated the work. They left because Google — a company with unlimited resources — could not figure out how to give them what they needed: compute, autonomy, speed, and a reason to believe the next model cycle would be different from the last.
OpenAI has two of them now. Anthropic has one. Cohere has one. Sakana AI has one. Inceptive has one. NEAR Protocol has one. Adept has one.
Google has zero.
And the model sitting at #7 on the Arena leaderboard is the proof.
Three Things to Watch
Gemini 3.5 Pro's release date. It was promised for June, then delayed to July. If it doesn't ship by August, the gap to Anthropic and OpenAI will be structurally insurmountable for 2026. Watch the date. It tells you everything.
The next resignation. Bloomberg reported that two additional researchers were in exit negotiations as of late June. If more Gemini core team members leave before the end of Q3, the model development pipeline is compromised — not because DeepMind lacks talent, but because institutional knowledge walks out the door with every departure.
Whether Google's engineers keep using Claude. When the people building Gemini choose to use a competitor's model for their own work, that is the most honest performance review you will ever read. Watch what they do, not what the benchmarks say.
I'll go first: I was a Gemini user. I switched to Claude three months ago. Not because Gemini is bad — it's not. It's fine. It's competent. It just isn't the best at anything anymore, and it used to be the best at everything.
That's what losing people looks like from the outside. It doesn't crash. It doesn't burn. It just... stops being special. And one day you realize you haven't opened the app in weeks.
Google is not dead. Google is not even close to dead. But Google is, for the first time in the AI era, clearly not the best. And the people who made it the best are now building the models that are beating it.
What's your take — is this recoverable, or is the talent loss structural? Drop it in the comments. Especially if you think I'm wrong. I'd love to be wrong about this one.
Written by a human who has been watching Google's AI journey since the Tensor Flow days. Reviewed by three other humans who use Claude. The AI that helped research this article has no memory of the people it's replacing. The opinions are mine. So is the discomfort of writing this on a browser powered by the company I'm writing about.
Top comments (1)
The talent exodus is the real story here, and it's worth separating from the benchmark table before responding to either. I can't verify several of the specific numbers in this piece, so I'm setting those aside. What I can speak to is the structural pattern, because it's one I've watched play out at smaller scale in security and infrastructure work for years.
The asset that walks out the door is never the tool. It's the person who understood why the tool was built that way. I've done enough architecture and security audits to know that documentation captures what a system does. It rarely captures why a specific tradeoff got made, or what almost broke during the build that never made it into a commit message. When that person leaves, the system doesn't fail immediately. It just stops improving the way it used to, and nobody can quite explain why for a while.
Compute reassignment as a signal is the sharper detail buried in here. Taking resources from someone who is supposedly co-leading your flagship effort tells them exactly where they stand, regardless of what the retention email says. That's not a talent problem. That's a governance problem wearing a talent problem's clothes.
Watching the Gemini 3.5 Pro ship date more than any leaderboard number. A missed date says more about organizational health than any benchmark score ever will.