TL;DR: Between September 30 and October 5, Google announced Gemini 4 Argon, which ties OpenAI’s GPT-6 Astra on Artificial Analysis’s Intelligence Index and posts the lowest hallucination rate of any top model. It also made Markdown a native file type in Drive and Docs. Put that next to Spark, which has become one of my favorite agentic tools, and free AI video in Google Vids, and I think the king is back. Google still has to open Argon to the public and fix how Gemini handles Office files.
If you wrote Google off this summer, you had reasons. On August 5, Demis Hassabis stepped out of day-to-day leadership at Google DeepMind, and Koray Kavukcuoglu took over Gemini. Jeff Dean announced his exit the same day. And the Gemini 3.5 Pro model Google had planned never shipped.
My own testing didn’t help. In August I ran my one-prompt Centipede test on Gemini 3.7 Flash. It beat three other frontier models, Claude Opus 5 among them, at building the game. Then it ignored my writing rules again and again, which matters a lot when your whole job is drafting to someone else’s spec.
Eight weeks after the shakeup, the picture looks different. On September 30, Google announced Gemini 4 Argon, its first model above the Flash tier in more than seven months. Five days later, it made Markdown a native file type in Drive and Docs.
The king is back. I think Google ends up on top of the AI pile, and once it’s there it’ll be very hard to knock off. That assumes Google can get out of its own way, which has never been a safe bet.
Spark puts Google ahead on agents, if your files are in Drive
Agents are where this race gets decided, and Google starts with the best seat. Gemini Spark, announced at I/O on May 19, is a personal agent that runs in Google’s cloud around the clock. It keeps working after you close your laptop, and it checks with you before high-stakes moves like spending money or sending email.
I’ll admit my bias. Spark is one of my very favorite tools, and I’ve put a ton of travel planning and research through it. It’s handled seasonal ski house rentals and multi-city flights across European cities. When I had social plans to sort out in different countries in Europe, Spark took those on too. And it’s tracked down specialty stores and unique T-shirts for me.
Two searches turned into a purchase. Spark researched large-breed dog harnesses for Magnus, my St. Bernard, and I bought one of its top recommendations. It was fantastic. The second was a new lens for my camera - it not only figured out the exact lens I needed - but if found a local shop I could purchase it at on the same day.
Spark started on Google’s top Ultra plan. In July it opened to $20-a-month AI Pro subscribers in the US, with up to 15 tasks running at once, and it has since reached more countries and select business users. In September it arrived in the new Gemini app for Windows.
Spark plugs straight into Gmail, Calendar, Drive, Docs, Sheets and Slides, because Google already holds your email and your files. Rivals like Claude Cowork and ChatGPT agent can get there too, but somebody has to set up the connections and permissions first.
For firms, Google says Spark in Gemini Enterprise runs in a managed sandbox on Google Cloud and needs explicit approval for risky actions like sending email. It can also reach Microsoft SharePoint and OneDrive through Gemini Enterprise connectors.
All of that assumes your files are in Google’s cloud, and a lot of the smaller firms I work with keep everything on a network shared drive. When they point a desktop agent like Claude Cowork at that drive, things slow to a crawl. We’ve had to change workflows and move active files to a local drive, which speeds things up dramatically.
Spark works on files already in Drive, so in principle it skips that bottleneck. I haven’t tested that yet. The king owns the firms already on Workspace. Everyone else pays a migration tax before Spark is even worth a test.
Argon puts Google back in the top tier
Gemini 4 Argon is the first model in Google’s Gemini 4 line, and Google’s launch post goes straight at my world. It names legal and finance work as targets and says Argon leads Harvey’s Legal Agent Benchmark for legal research and drafting. Google also claims the top spot on the Vals Index, which weights finance, coding, legal and tax work by each sector’s share of US GDP.
Vendor charts are marketing until someone else checks them. Artificial Analysis checked, and scored Argon at 53 on its Intelligence Index. That ties GPT-6 Astra and lands 23 points above Google’s previous non-Flash model, Gemini 3.1 Pro Preview. Their verdict was that Google is back among the top three labs.
Argon isn’t first. Claude Opus 5.5 leads the same index at 58, and even Google’s own chart shows Astra winning a few tests.
Two specs matter for long legal work. Google raised Argon’s output limit to 1 million tokens, up from 64,000 on earlier models. Intro pricing is $2 per million input tokens and $10 per million output, and after the promotion it goes to $4 and $20.
One catch with that price. Argon thinks long. Artificial Analysis measured about 62,000 output tokens per task against 27,000 for Astra, so at full price each task costs roughly 1.2 times what Astra does.
So my bet rests on distribution. When the top models sit within a few points of each other, the company already inside Gmail, Chrome, Android and Search wins on reach.
The number lawyers should care about
I’d tape this number to every managing partner’s monitor. On Artificial Analysis’s AA-Omniscience test, Argon’s hallucination rate is 15%. GPT-6 Astra comes in at 51% and GPT-6.1 Sol at 54%. Artificial Analysis says that’s the lowest of any model scoring 45 or higher on its index.
The test looks at what a model does when it doesn’t know the answer. A low rate means Argon usually admits it’s stuck. Astra guesses wrong about half the time it doesn’t know.
There’s a trade. Argon answered 50% of the questions correctly, against 63% for Astra. It knows less, and it’s more honest about the gaps.
For lawyers, I’ll take that trade every time. An honest “I don’t know” costs you one more research step. A confident fake citation can end in a sanctions order and a call to your malpractice carrier.
One caution before anyone gets excited. AA-Omniscience is a broad factual-knowledge test, so a low score doesn’t replace cite-checking. And I can’t test Argon on legal work yet. Outside Google’s early-access testers, nobody can.
Google still has to win the file fight
Since April, Gemini has been able to generate Word and Excel files from a prompt, one file at a time, and PowerPoint isn’t on Google’s list at all. For a law firm, that’s the wrong end of the problem. The file that matters is a Word document carrying opposing counsel’s tracked changes.
On October 5, Google made Markdown files native in Drive and Docs. You can preview a .md file in Drive, then edit and comment on it in Docs, and it stays a .md file. Google’s own announcement admits the old import “often altered formatting, stripped comments, and fragmented files.”
Markdown matters because it’s the format AI tools write in. Google’s Chandu Thota called it “the common language between humans and AI agents.” The change reaches everyone over about 15 days, with no admin switch to turn it off.
Google needs to get its act together and give Office files the same treatment. That means Gemini working inside a client’s .docx, redlines and all, without converting it. Until then, firms that run on Microsoft 365 have little reason to move. Spark’s SharePoint and OneDrive connectors are a start, and Google needs to finish the job. If there’s one thing I really want Google to address - it’s this.
Skills replace Gems, and projects are still missing
On September 30, Google said skills will replace Gems, which business customers lose in March 2027. A skill is a saved set of instructions you call up with a slash, and you can stack several on one task.
Write one skill for your citation format and another for a client’s defined terms, then run both on the same draft. A skill won’t fix a model that ignores instructions, but it makes good instructions easy to reuse.
Skills don’t give you a real projects workspace. ChatGPT and Claude both let you build a project with its own files and instructions. Every chat about that matter lives there too. Google’s closest answer is notebooks, which sync with Gemini Notebook, the tool it used to call NotebookLM. It feels like a research tool bolted onto a chatbot. Google needs a true projects approach, and I’d put that near the top of its list. Until then, Google can sit inside the inbox and still lose the matter, because the matter lives in a project and Google doesn’t have a real one yet.
Video, and the model I’m waiting for
Google has replaced Veo with Gemini Omni in the Gemini app, and since September 23 it gives Omni 1.1 Flash away free in Google Vids to anyone with a Google account. In March, OpenAI announced it was shutting down Sora, which takes Google’s best-known video rival out of the race. What Google hasn’t shipped is a Pro-class Omni, and if one lands soon I don’t see who catches it in video, for consumers or for business.
How Google could still blow it
You can’t use Argon today. Google is limiting it to cyber defenders in its Fairwind Program, with paid API customers and Ultra subscribers next and no date attached.
Reach doesn’t guarantee use either. Early this year, Statcounter had Gemini at about 7.2% of AI chatbot interactions against ChatGPT’s 80.5%.
My August test still nags at me too. A model that ignores a style guide is likely to ignore a court’s formatting rules, and nobody outside Google can check yet whether Argon fixed that.
The fair objection is that leaderboards move weekly. In late September, OpenAI replaced GPT-6 Sol after just seven days. I’ll grant that a 53 could look like a 51 by Christmas.
My bet holds anyway, because the model was the one piece Google lacked. It already had the inbox, the browser, the phone and the documents. Now it has a top-tier model that admits when it doesn’t know.
What to do Monday
Run a two-week Spark test on non-client work. Start with the kind of errands I use it for, like travel or research, then give it one recurring office job such as a Monday inbox recap. Keep client data out until your admin has reviewed Spark’s business controls.
Add “I don’t know” questions to your model tests. Ask each tool three questions you know have no settled answer under your state’s law, and score whether it admits that. Rerun them on Argon the week it opens to Ultra subscribers.
Keep any AI contract you sign this quarter to 12 months. With Argon still gated and OpenAI swapping models in seven days, a three-year commitment is a bet you don’t need to make.
Google finally has a model to match its reach. I’d bet on the king.
Thank you for reading. If you enjoyed this, please share this with others.
Here are two shots of Magnus doing his best to stay out of this insane heatwave impacting Southern California.
