The June slide said the flagship would ship that month. The July post said partners were testing it. August took the chief scientist, the Gemini co-lead, and the DeepMind CEO’s operating job. The playlist from the old Research all-hands died with the corp login.
By Ryan Merket · Published
Primary source: x.com
Why it matters #
Google's response to a missed flagship deadline has been organizational and commercial: put an operator over Gemini, ship cheaper Flash models faster, and back the marquee researchers leaving to build outside Google.
On August 14th, nine days after he left Google, Jeff Dean went looking for a playlist.
He had posted it in 2020 for a Google Research all-hands. Colleagues had made it as a joke and a gift: “Jeff’s C++ Bangers.” The link was dead.
This used to be a playlist some of my colleagues made for a Google Research all hands, but the playlist itself has disappeared with my
[https://www.google.com/]account shutdown. Someone asked if he still had a copy. He did not. He built a new public list from songs he had Liked in YouTube Music and left it at that.
Dean was Google’s 30th employee. He spent 27 years there. MapReduce, Bigtable, TensorFlow, the TPU program, Google Brain, then chief scientist of Google DeepMind. On August 5th he posted that he was founding a public-benefit corporation called Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. Mission line: automate machine learning, science, and engineering. Alphabet is a founding investor and Cloud partner. Sundar Pichai’s memo that morning called it an “incredible 27-year run” and said Dean was “at a moment where he wants to try something new.”
The corp account still closed hard enough to take the music.
That is the summer, if you only read what Google and its people put on the internet.
The date they wrote down
Alphabet’s June 2026 investor presentation is still up. Gemini 3.5 Flash is now available, and we expect Gemini 3.5 Pro will be coming in June.
Two weeks earlier, at I/O, Demis Hassabis posted the Flash launch himself. Frontier-ish numbers. Four times faster than other frontier models. Cheaper.
Gemini 3.5 Flash is amazing!
- Performs better than 3.1 Pro on coding & agentic tasks
- 4x faster than other frontier models
- 12x faster in @antigravity - 800 tokens/sec!
- Often at less than half the cost And Pro to come…
Noam Shazeer, Gemini co-lead, was on the same launch thread talking agentic workflows and Terminal-Bench. June arrived. Pro did not.
On June 18th Shazeer posted that he was joining OpenAI.
I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.
It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to work with all of you.
His bio now says Engineer. Google had spent on the order of $2.7 billion in 2024, mostly as a Character.AI license, to bring him back after he walked out the first time. He lasted less than two years.
“Soon”
July 21st, Logan Kilpatrick, the public face of the Gemini API, posted two things into the same afternoon.
One:
We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress : )
(The post) Two:
3.5 pro is testing with partners! will hopefully land soon.
(The post. He said a version of it twice.) The next day Pichai got on the Q2 call. Prepared remarks, later posted on blog.google: new Flash variants, a Cyber model paired with CodeMender, Gemma past 900 million downloads. Then the sentence the June slide had been waiting for.
Gemini 3.5 Pro is currently in testing.
Then the swing:
We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress we are seeing at the frontier.
He also said the company was supply-constrained. Model APIs were processing about 22 billion tokens a minute, up from 16 billion the quarter before. More than nine million developers a month. Gemini Enterprise in 90 percent of the Fortune 100. Cloud revenue up 82 percent. On X after the print he called it an amazing quarter.
Analysts asked about the frontier anyway. In the Q&A he said there were attributes where Google was still there and an area where it was not. Coding. Agentic coding. He said they needed to improve. He talked about a faster cadence, Flash-class models on something close to a monthly clock, and a larger base model for 4.
The stock had already started to give back the year. The miss on a dated flagship is the part of an earnings call that does not fit in the token chart.
The same Wednesday
August 5th, blog.google: [“The next chapter of our AI momentum.”](https://blog.google/company-news/inside-google/message-ceo/next-chapter-ai-momentum/)
Pichai: 950 million monthly users on the Gemini app. Six days later he [posted one billion](https://x.com/sundarpichai/status/2087222656819241292) and called it Google’s fastest-growing product, the 14th to that mark. Flash in demand. Cyber live.
We are committed to being at the frontier, and are super focused on the areas where we need to improve.
Demis Hassabis off day-to-day DeepMind. New titles: Chair of Google DeepMind, Chief Scientist of Alphabet, still CEO of Isomorphic Labs. Pichai wrote that Hassabis had described the company as standing in the “foothills of the singularity,” and that the two of them had been talking about a job that let Hassabis put his attention on “actively shaping the future of AGI.”
Hassabis’s own note is the one that sounds like him.
I’ve been working towards AGI my whole life and now, like many of you, I feel it is close at hand. It’s critical that we collectively get the next steps right to ensure this all goes well for humanity and we usher in an incredible new age of discovery and wonder.
With this backdrop, I’ve decided that now is the right time for me to hand over my day-to-day operational responsibilities at GDM, so that I have the time and space to focus on the big picture.
He posted a shorter version on X. Koray Kavukcuoglu, DeepMind’s CTO and Google’s chief AI architect, would run the lab as SVP, reporting to Pichai, owning Gemini models, frontier research, the Gemini app, and the developer teams. Hassabis would advise from London, lean into Isomorphic, and repeat the line he always repeats: the number one application of AI should be human health. Cure cancer.
The Gemini models are in good hands with Koray and the leads, as they have been for a while, and I’m excited about the great progress we’re making with our new models including Gemini 4.
Same morning, Dean’s Discovery Loop post. Same morning, Oriol Vinyals, Gemini co-lead, thirteen years after a cold email from Dean brought him to Google:
Thirteen years ago, a quick email exchange with @JeffDean brought me to @Google. After an unforgettable journey from Brain, to DeepMind, and to @GoogleDeepMind, working on seq2seq or distillation, and co-leading huge projects such as AlphaStar, and Gemini, I’m saying goodbye to GDM.
The best part? The same person who brought me in is challenging me to take the next leap.
Excited to build Discovery Loop (@DiscoLoopAI) alongside @JeffDean, @quocleix, and @Sanjay_Ghemawat
Deeply grateful to everyone who made GDM feel like home. Time to try something new and to keep it drastic! ✨♾
His X bio still reads VP of Research & Deep Learning Lead, Google DeepMind. Gemini co-lead. The title did not update with the goodbye.
Wired later ran Dean and Vinyals on the record about why a four-person company. Dean: cloud compute now lets a small team skip building the warehouse. Vinyals: “In a large organization there is always a lot of inertia you have to overcome to make any radical changes. We want to build something different.” Dean told a Stanford audience the Gemini team had a plan to make the models “awesome,” then talked about focus. After he left he also said, in public, that coding had been a conscious lag. The lab had wanted a model that was good at many things. Programming waited.
That is as close as the people who built Gemini have come to a diagnosis, and they said it after the badge stopped working.
What they shipped instead
August 13th, Koray’s first real post as the person who owns the models. Gemini 3.7 Flash. Introductory price at half of 3.6 Flash. Three months from 3.5 to 3.7. He put the coding numbers in the tweet.
Today we're launching Gemini 3.7 Flash - our latest workhorse model for coding and agentic workflows, with an introductory price at half the original cost of 3.6 Flash.
We have been iterating rapidly with the Flash series, going from 3.5 to 3.7 in just 3 months, making it more helpful across a wide range of tasks:
- Software Engineering (DeepSWE v1.1): 37.0% ➔ 65.3%
- Web Development (Code Arena Elo): 1506 ➔ 1588
- Enterprise Automation (AutomationBench): 13.4% ➔ 30.4%
Pichai quote-tweeted ARC-AGI scores and said 3.7 Flash had smashed prior Gemini growth records in week one, then landed in Search and the Gemini app.
Josh Woodward, who runs the consumer app, had already published the other ledger. In July he ranked 1,400 user replies. Number one: Workspace integrations. Number two: tool calling. Then Projects, MCPs, Deep Research, watermarks, dictation, mobile scrolling, celebrity guardrails. In August he came back and marked items done or in flight. “3.7 Flash showed improvements in tool calling, more coming here.”
That is the product org saying the assistant that just crossed a billion users still drops tools. They said it themselves.
The rest of late August is the portfolio Google can actually print on a calendar. Gemini 3.5 Transcribe, 85-plus languages, Pichai on the video. Omni Flash 1.1 for video. Robotics 2 in July, Hassabis congratulating the team for robots that tie knots. Pixel 11 with “Gemini Intelligence.” None of those posts mention 3.5 Pro.
Kilpatrick’s July 21st “soon” is still the last dated public sentence on the flagship.
The rule on the wall
Sergey Brin told the All-In podcast a story that explains more of the coding hole than another benchmark screenshot.
Google kept an internal list of tools engineers were allowed to use to write code. Gemini was on the no list. Brin said he found it, fought it for weeks, could not kill it, and took it to Pichai. Nobody was really enforcing it, he said. It was still on the page. Weird historical reasons. Fear that proprietary code would leak into training. The Times of India ran the remarks in July. Brin said them on tape.
A company trying to win software agents had, for a stretch, told its own software agents they were not a permitted way to write software. Then Pichai told the Street they were behind on exactly that.
You do not need a source in the building for the rest of the shape. The June deck dated Pro. The July API account slipped to “hopefully.” The August memo split the lab so the AGI-and-cancer job and the ship-Flash job had different bosses. The people who had been listed as Gemini’s technical co-leads left the same morning, and Google bought a stake in the thing they left to build. Dean’s login died and took a Research all-hands playlist with it.
Hassabis wrote that Google is the only company with the full stack and that he firmly believes they will lead. Pichai keeps pointing at tokens, Cloud, Search, YouTube, and a billion people in the app. Those numbers are real. They are also the numbers you point at when the named Pro model is still a partner-test.
What this is
It is not a secret that large companies are slow. Google has been writing memos about that since before ChatGPT. The public record this summer is more specific. They put a month on a flagship. They missed it. They said the miss was coding. They replaced the miss with cheaper models on a faster clock and a training run they will not date. They moved the scientist who talks about singularities into a chair, and they put a Mountain View operator on the release train. The builders of the last leap incorporated a four-person company whose entire pitch is that a small group can now rent the computers and skip the org.
Dean did not write a thread about any of that. He lost a playlist and made another one.
3.5 Pro is still “soon.”
How this story was made
Assigned and edited by a human. Drawn from company blogs, an investor deck, an earnings transcript, and named posts on X. No anonymous sourcing. No request for comment. Google’s public statements are quoted as they stand.