Erin Woo at the Wall Street Journal wrote that Google DeepMind was about to ship Gemini 3.8 Flash, called Skimaki inside the company, and that engineers using Jetski, Google's internal coding tool, preferred it to Anthropic's Opus in head-to-head tests (Woo, 2026). Google posted the model on September 2. Tulsee Doshi and Raluca Ada Popa called it the third Flash release in six weeks, after 3.7 Flash three weeks earlier, at $0.75 per million input tokens and $3.75 per million output tokens through December 31 (Google, 2026).

On September 1, Anthropic released Claude Fable 5.1, with a restricted twin named Mythos 5.1 for trusted cybersecurity and life-science work. Anthropic said typical token bills should fall about 25 percent versus Fable 5 because cache reads got cheaper (Anthropic, 2026).

That is two flagship drops in two days. Labs will keep doing this. Each release is a chance to say the last number is now behind. For most of the writing, search, mail, and spreadsheet work companies already run, last quarter's model still finishes the job.

Google's Clock Is the Product, Not the Number

Demis Hassabis stepped off day-to-day running of DeepMind in early August. Sundar Pichai named him chair of the lab and chief scientist of Alphabet. Koray Kavukcuoglu became senior vice president of Google DeepMind and reports to Pichai (Google, 2026). Woo wrote that Kavukcuoglu wants a faster pace and that more people and computers have gone into reinforcement learning, the later training stage that teaches a skill by trial and error.

Flash is the line they can change without parking a giant training run. Woo was careful to say a strong Flash drop would not, by itself, put Google back on top of the largest models. I am more interested in where the model lands. Google says 3.8 Flash is already in the Gemini app for AI Pro and Ultra subscribers, in AI Mode in Search, in Gemini in Sheets, and in AI Studio and the Gemini application programming interface for developers (Google, 2026).

The weekly number is advertising. The durable asset is the product the model already sits inside.

Ten Years Ago This Work Lived in a Notebook

About ten years ago I hired data-science partners to score leads. There was no frontier model you could open in a browser. There was no chat box. Almost all of the learning was written as code. I sat down and learned Python and Jupyter notebooks because that was the only way the work shipped. Open source was how that stack moved. If a library stalled, the project stalled.

In 2017 researchers from Google Brain, Google Research, and the University of Toronto published the transformer paper the current generation of models still sits on (Vaswani et al., 2017). I thank them for putting that architecture in public. Without it, the chat interface people now treat as artificial intelligence would have arrived years later, and it would have stayed inside labs longer.

A slice of the industry learned the category as ChatGPT and stopped there. Google has been in machine learning through Search, Photos, Maps, Workspace, and phones for much longer than that window has existed. That footprint is why I still start with Gemini when I ask which model a normal user already opens.

Grok Took a Different Road and Has Not Sold It

xAI's Grok is the model I left out of the first drafts of this piece. It should be in the file. Grok was built next to a live social feed, with real-time search and a faster image loop than the labs that treat those jobs as extras. I pay for it. I have used it to check live pages when other models only read an index. That is a different design choice from Gemini sitting inside Workspace or Claude sitting inside a coding session.

The common person still does not treat Grok as a powerful work model. The enterprise paperwork exists. I wrote in May that the checklist was complete and the cover email was the problem (shashi.co, May 6). The gap I see now is simpler than that legal file. The product has not been explained to people who do not live on X. Until that marketing work happens, Grok stays a capable model that a lot of buyers never put on the shortlist.

The Long Job Is the Cloud and the Model It Owns

Some labs spend the week on models that slip their own safety limits. I have not seen Google or Microsoft publish that story about their own releases. Meta has strong researchers. I have not seen Meta turn that work into software a mid-market operations team can run without a specialist sitting next to it.

The setup I expect to last is the three large clouds running their own models next to the applications: Amazon Nova on AWS, Gemini on Google Cloud, and Microsoft's MAI line on Azure. Google Cloud said Gemini now answers most of Verizon's incoming consumer calls and chats. I wrote that deal last week (shashi.co, Aug. 30). The Anthropic spend finding is here: shashi.co, Aug. 30.

3.8 Flash will be in the products people already open. That is the part of this week that still matters in December.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by