Gemini 4 Argon は何に向いているのか

Two weeks after being announced, Google's new Gemini 4 Argon model is here, available for testing, and the questions abound. Will it live up to its high benchmarks? Can Google finally compete when it comes to agentic coding? If not agentic coding, are there other use cases where Google might now be in the lead? Today we're looking at the newest frontier model we have access to, Gemini 4 Argon, and where it fits. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Section, Robots and Pencils, and Blitzy. To get an ad-free version of the show, go to patreon.com slash AIDailybrief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at sponsors at AIDailybrief.ai. And to find out about everything happening in the broader AIDB ecosystem, go check out the website. You can find access to our past episodes, each of which is divided into a million small sharing nuggets to make it easy for you to share with friends and colleagues.

You can also get access to our free training programs. Or if you want to go deeper, you can check out the paid training programs that are offered via Superintelligent, like the new AI Champions program that we just announced. Last note today, it's not completely essential, but there are some benefits to consuming this one as a video, given that a lot of the demos of Gemini 4 Argonne are video. The episode will be in video on both YouTube and Spotify, and video will be coming to Apple soon as well. We kick off today with some palace intrigue. On Tuesday afternoon, the OpenAI Newsroom account posted, We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI and were grateful to the agents and prosecutors who moved quickly to make an arrest. TLDR, an OpenAI employee has been arrested and charged after a listening device was discovered in an executive conference room. According to a DOJ complaint, around two weeks ago, OpenAI discovered an audio recording device disguised in toilet paper

and plugged into the back of a TV in a conference room at their San Francisco offices. As you might imagine, meetings in that room routinely involved discussion of highly confidential technical, commercial, and strategic information. As would come out later, security footage showed employee Abdulrahman Al-Fozan near the conference room on multiple occasions prior to the discovery of the device. Abdul Rahman had no authority to access the room or participate in the meetings held there. Alphazan had been employed by OpenAI for around a year as an engineering lead working on agentic products. He had previously worked for Meta for five years. According to the complaint, Alphazan used a productivity tool to make inquiries about OpenAI's meeting room layout, executive locations, and details of meetings with external partners. A chatbot owned by Alphazan's spouse was also used for queries related to building a continuous audio recorder, customizing Wi-Fi connectivity, improving audio and live listening, and encrypting the SD card, among other things. The account was also used for queries about the specific device that was found, identified by its customer header and unique MAC address. These queries all took place in the week prior to the device being found. Finally, on October 5th, Alphazon entered the conference room

and checked behind the TV, a moment that was captured on camera. The same day, upon searching an unattended bag, OpenAI security personnel found Alphazon's wallet and ID, as well as a second recording device of a similar make. Al-Fazan has been charged with attempting to use an electronic device to intercept oral communications, and the FBI is conducting further investigations. If this were just corporate espionage, there would be plenty to keep all of us in the peanut gallery interested. What takes this case to the next level, however, is Al-Fazan's status as a Saudi national, the suspicion being that this is not only corporate espionage, but straight-up international espionage. The DOJ announcement was overt in calling this a potential national security incident. FBI Director Cash Patel said, According to the complaint unsealed today, Abdelrahman Al-Fazan allegedly concealed a covert recording device inside his employer's executive conference room to secretly capture sensitive communications about the company's superintelligence technology. Protecting American SI innovation is essential to safeguarding our nation's economic strength and national security, and the FBI views that mission as one of our highest counterintelligence priorities.

Patel noted that National Intelligence Director Jay Clayton has been working closely on the case, Clayton, of course, also being the newly appointed head of the administration's AI task force. U.S. Attorney Craig H. Misakian added, The Bay Area and Silicon Valley lead the world in developing tomorrow's technology, and our office will continue to stand at the forefront of protecting that American innovation. For us, it is a matter of national security. The Bay Area in general and the AI industry in specific have long been a target of international espionage. What makes this somewhat different is Saudi Arabia, nominally a U.S. ally, being implicated. Journalists on X got to work unpacking the backstory. Scott Stedman from Newsground said that, according to his sources, Al-Fazan was the only Saudi employee at OpenAI. Adding another piece of red string to the corkboard, Stedman noted that Al-Fazan is also a venture partner at RAED Ventures, where he works alongside Prince Khalid bin Bader al-Sad. Stedman identified Prince Khalid's father as a member of Saudi intelligence, but officially he's currently the ambassador to the UK. Holding aside the specifics, essentially the point here is that Al-Fazan has ample ties to the House of Saud and the Saudi government to ground suspicion.

As far as implications go, OpenAI claims their operations are secure. They wrote that, At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. Still, Chris McGuire from the Council on Foreign Relations believes, This is an extremely bad look for Saudi Arabia, and if the Saudi government is indeed involved, this should have significant implications for the U.S.-Saudi AI relationship. McGuire presented a lengthy analysis that's actually worth reading in full if you are interested, but to hit the high points, McGuire believes that AI labs will need to harden their security and presume they are already the target of less ham-handed, his words, espionage attempts. For Saudi-U.S. relations, he believes this is a very bad sign. The investigation is being handled by the DOJ's counterintelligence division, not the criminal division who would handle regular commercial espionage cases. McGuire wrote, If Al-Tazan was acting in coordination with the Saudi government, I do not see how any company could possibly host a frontier AI model in Saudi Arabia after this incident. Sadly, it is way beyond the scope of this particular show, but there is a lot going on right now in the U.S.-Saudi relationship, And while there are plenty of explanations for this that don't implicate the Saudi government,

it is an incident that could have some fairly significant implications for global AI politics. Moving over to the much more banal reality of our new agentic era, JPMorgan plans to block third-party agents from accessing customer accounts. During Tuesday's earnings call, agentic risk was top of mind for JPMorgan executives. CEO Jamie Dimon said, We're not going to allow just any outside agent to come inside the company, because it's just too risky for clients and for JPMorgan. You don't want rogue agents, and it could be thousands of them. While diamonds seem to be conflating the scary headlines around hugging face, with consumer agents like Muse and Instinct, the bank is still changing its policy. CFO Jeremy Barnum said, To the extent that we're able to block those, we are blocking them to protect customers and to protect ourselves. After the call, a spokesperson told CNBC, Protecting our customers is our top priority. Because it is not safe, we do not permit third parties to access accounts by obtaining customers' banking credentials and logging in on their behalf. We support secure, customer-authorized data sharing through open banking providers and are committed to enabling agenda connections securely.

This means that agents will likely at some point be able to access account information for basic read-only tasks like making a budget or finding old subscriptions to cancel, but JPMorgan customers are unlikely to be able to get their agents to pay bills or settle invoices. Diamond said frankly that customers don't understand the risks they're running, commenting, If people are using them they should understand the risk They should understand that there a liability transfer Essentially the bank won indemnify faulty agents in the same way they do for fraud Now agentic banking has been a touchy subject since the introduction of Muse brought personal agents to the mainstream Late last month Apollo chief economist Torsten Slock warned that agents seeking a higher yield on savings account for their users could trigger a bank run. If one puts much credence in that scenario, J.P. Morgan would be ground zero with an average savings account rate of 0.01%. J.P. Morgan doesn't seem concerned about that particular outcome, but rather the potential liability of personal agents spending without proper authorization. Basically, with this policy change, they've made it clear that they won't be refunding customers if Muse goes on a spending spree. AI for Finance founder Bob Su thinks that while this was completely inevitable, it's

also not the right approach to a real problem. Bob wrote, Banks blocking AI agents is the most predictable move of 2026. They're right about the risk, wrong about the fix. Users don't need agents locked out. They need agents locked down. Scoped permissions, hard limits, a kill switch. Meanwhile, on the other end of the spectrum, Shopify has thrown open their doors to agents of all stripes. Last month, Shopify partnered with Muse to allow agents to browse and shop across their platform, and the key technical detail was that Shopify would provide API access directly to their back end. This was a mutually beneficial arrangement. It allowed Muse to produce better results, while allowing API access meant Shopify wasn't paying web hosting costs associated with agents scraping their front end. Now, Shopify has extended this arrangement to agents from any provider, Users can now connect their agent via the Shopify app for a fully supported agentic shopping interface, regardless of whether they're using Muse, Town, Hermes, Instinct, Dots, or any of the other dozens of personal agents on offer. The move is part of a broader trend that's seeing two diverging approaches to servicing agents. Some, like Amazon and eBay, and JP Morgan as we just heard,

are locking their platforms down and forcing customers to use their proprietary agents. That approach maintains control of the customer relationship and associated data, but it risks driving agentic customers to other platforms. The other option is to facilitate agentic interactions directly by offering a headless interface for agents, as Shopify has done. The same strategic divergence is playing out in SaaS, where platforms like Salesforce have leaned into headless product by opening their APIs to third-party agents. SAP, on the other hand, has tightened controls on APIs, securing their moat at the cost of customer choice. This schism between open and closed platforms is one of the key battlegrounds as agent adoption grows, and it's still very unclear which strategy will win out. Lastly today, Character AI has filed for bankruptcy, marking the end of the road for a once popular consumer AI product. Back in 2024, the olden days, Character AI was the first breakout hit beyond ChatGPT, ranking at third place in A16Z's list of top consumer AI apps. It allowed customers to create their own character-based roleplay using chatbots. Many assumed that when Google acquired most of the company's top talent

in a multi-billion dollar acqui-hire deal, that Character would be living on borrowed time, but for a while, the company maintained a dedicated following. The bankruptcy tells two stories. First, there's the incredible legal liability from playing fast and loose with technology and business decisions. Beyond lawsuits over customer self-harm and chatbots mimicking doctors, Character AI ran afoul of Disney in 2025 after adding support for unlicensed characters. The bankruptcy filing states that Character AI has between $50 and $100 million in liabilities that they're seeking to discharge. Disney has emerged as the stocking horse bidder to salvage assets from the bankruptcy. They've agreed to pay $20 million for certain intellectual property, and recently hired Character AI's CEO to facilitate the technology transfer. Beyond that, there's a broader question about whether a consumer-focused business model can function in 2026. 404 Media, who it should be known, is pretty wildly antagonistic to AI and tech in general, made the bold claim that the economics of AI don't work based on the bankruptcy filings. According to an affidavit from CEO Colin Adams, Character AI was spending $10 million per month on compute, with peak monthly subscription revenue of $2.8 million.

Other months saw less revenue, even dipping below $1 million a month at times. While 404 presented this as a condemnation of the AI business model as a whole, I think the more interesting questions, which don't need to apply to absolutely everything to be interesting, are about what consumers are going to be willing to pay for and how much they're going to be willing to pay, and how that does or doesn't line up with products on offer. Discovering that people aren't willing to pay enough for a specific implementation of roleplay chatbots doesn't mean that there's no business model in consumer AI. It means that people didn't want to spend enough to justify that particular set of roleplaying chatbots. Cleometric summed it up, a strange company. They got addicted to free user analytics numbers, dragged their feet on moderation for adult content, and got stuck in this will-they-won't-they loop. Half measures all around, they missed the boat one too many times. Wrote computer scientist John Kempman, this is the pets.com of AI. People genuinely do not understand the AI industry. I can't believe people genuinely think that character AI filing for bankruptcy means Anthropic is next. Still an interesting moment as the early phases of AI cede to something different.

For now, however, that is going to do it for today's headlines. Next up, the main episode. If you're leading AI inside an enterprise, you already know that the gap right now isn't capability but execution. That's why KPMG's You Can With AI is back with a new season featuring conversations with leaders like Sarojit Chatterjee of Emma, May Habib of Reiter, McKesson CIO Ellery Fisher, and others focused on practical execution. What's working, what's not, and what it actually takes to move from pilots to real scaled impact across strategy, data readiness, governance, workforce, and value. And of course, it's co-hosted by me, Nathaniel Whittemore. Go listen and subscribe at www.kpmg.us slash AI podcasts. That's www.kpmg.us slash AI podcasts. Here's a harsh truth. Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized. Half of companies have AI tools, but only 12% use them for business value. Most employees are still using AI to summarize meeting notes.

If you're the one responsible for AI adoption at your company, you need Section. Section is a platform that helps you manage AI transformation across your entire organization. It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value. The result? You go from rolling out tools to driving measurable AI value. Your employees move from meeting summaries to solving actual business problems, and you can prove the ROI. Stop guessing if your AI investment is working. Check out Section at sectionai.com. That's S-E-C-T-I-O-N-A-I dot com. The best teams don't have a single star carrying everyone else. They know their own strengths and each other's weaknesses and play to both. That's the team Robots and Pencils has built on purpose. Nobody there is grinding through busy work to pad a headcount number. People come for the hard problems and they stay because everyone around them is leveling up at the same time. In a market full of companies that are just trying to hire fast, that's worth a look. Check out robotsandpencils.com slash careers. Here's why most legacy modernization projects fail.

The AI doing the work can't understand code bases at scale. It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem. Blitzy solves this the way it solves everything. Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph. Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head. From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith to microservices transformations, all validated end-to-end. One Blitzy customer modernized the $10 million monolithic insurance stack in 16 weeks against a 137 baseline with coding agents That 9x compression Retire technical debt while accelerating your roadmap See how at Blitzy That B dot com Welcome back to the AI Daily Grief. Today, for the second day in a row, we are looking at a big new model release. And what's interesting is that just like yesterday,

this is a model release from a company that has, for at least the last phase of AI, been considered a little bit behind the eight ball. I'm talking, of course, about Google's new Gemini 4 Argon. And while the perception of Google has been nowhere near as bad as it was for, for example, Meta earlier this year, it's been a bit of a roller coaster where Google was far behind, only to then launch themselves right back into contention, spending most of 2025 right on par, more or less, with OpenAI and Anthropic, only to almost completely miss the 2026 wave of agents and agentic coding. Gemini 4 Argon was announced a couple of weeks ago at the end of September, but it wasn't available yet, which many took as not such a good sign. Still, for the very many of us who were quite willing to give Google the benefit of the doubt, and not at all comfortable with writing them off, it was really always just a waiting game to see what the model could actually do when we could get our hands on it. As we dig in, it's important to note that I believe that the right question about any new frontier model has shifted. Whereas it might have been before, how good is it? with the implication being, should I switch?

The question is now much more, where does it fit? In other words, what things does this particular model do well or different that should make me consider integrating it into my overall model stack? So that's what we're going to explore today. To recap the basics from a couple of weeks ago, this is a brand new model built on a brand new pre-train after Google scuttled the Gemini 3.5 Pro release earlier this year. It's unclear exactly how large the model is, but Google is presenting it as an ultra-large model in the same class as Astra and Fable. To recap the benchmarks that we got a preview of, at least by the benchmarks, Argon appears to be broadly a match for frontier models from the rival labs. For knowledge work, Argon is state-of-the-art on the VALS Index, Automation Bench, VALS Finance Agent, and Harvey's Legal Agent benchmark. And other than the core VALS Index, each of these benchmarks have a wide gap above competitors. On Harvey's Legal benchmark, for example, Argon is 13 points ahead of its closest competitor, Fable 5.1. On coding, the results are more mixed. Argon is state-of-the-art on DeepSuite, scoring a few points above Opus 5.5,

but on FrontierSuite, Argon is behind the three major rivals, and about 10 points off of the leader Astra score. Similarly, on TerminalBench 4.0, Argon is 9 points behind Opus 5.5, who has the leading score. Although to be fair, that presentation is probably a little misleading, the spread of coding scores has widened with recent updates to the benchmark, and although they are behind, Argon is firmly in the Frontier range, just not at the head of the pack. Google highlighted a variety of other benchmarks across agentic science, cybersecurity, and multimodal, with Argon state-of-the-art across each category based on the benchmark selected by Google. One other benchmark worth highlighting, given that a lot of the questions are going to be about agents, is Argon's score on OSworld, which is the standard computer use benchmark. Argon scored 69.2%, which is just shy of Astro's 72.6%, which when announced was considered a paradigm shift in AI computer use. Independent benchmarking from artificial analysis similarly placed Argon squarely at the Frontier. It scored 53 overall, putting it tied for fourth place with Fable 5-1 and GPT-6 Astra. Muse Wonder Full and Sonnet 5-5 are tied for second with 56 points,

while Opus 5-5 is in the lead with 58 points. Realistically, and so we don't get too hung up on the exact numbers, there is now a group of between six and eight models across four US labs with a legitimate claim to being competitive at the Frontier. depending, of course, how you feel about GPT-6-1-SOL and Grok 4.7. But of course, we no longer live in just the world of performance, we live in the world of efficient performance. And cost is a little more difficult to dissect due to the way that Google has rolled this model out. On the API, Argon is available with a 50% cost reduction until November 30th. That allowed the model to complete the AA benchmark run at $199 per task, very similar to the cost of GLM-5-3 and significantly cheaper than all Frontier Rivals other than 6-1-SOL. Once the subsidy goes away, Argon will be more expensive than Astra, but still around two-thirds the price of the Anthropic models. Trying to sum it up, going purely from the numbers, Argon looks like a frontier AI at a discount, which presumably would put Google back in the game. And yet, for many, the biggest question to be answered by the release of Argon is, can it code?

Coding has been Google's Achilles heel over the past year, and there's been speculation that poor coding performance led Gemini 3.5 Pro to be shelved over the summer. Pavel Hurin ran Argon through his bug-hunting task and found it to be... inadequate? It performed about as well as Mew Spark 1.3 and significantly behind Astrofable and Opus. Hurin still gave Google credit for making progress, writing, A way better model than Gemini 3.8 Flash. Looks like Google is back. Noah Dunnigan from Railway was less impressed, posting, This model is useless when it comes to real-world code. I'm not mad, just disappointed. This, however, actually kind of makes sense for Google. I think they have given up the programming model race and are instead aiming for knowledge workers. Others found similarly not so great results. Developer, content creator, and entrepreneur Theo wrote, Just ran Argon on my code-based analysis bench. Not looking good. Only model with a greater than 50% false claim rate. Dan Shipper from Every writes, Gemini 4 Argon really struggles on our senior engineer benchmark test. Comes in at 21%, far below Sol 6 and even Muse.

Analyst Max Weinbach did respond to Dan saying, it really depends on the harness you're using. I didn't love it in anti-gravity, but found it was really good in Cloud Code and Open Code. However, as some pointed out, using the model and other harnesses is technically a violation of the terms of service. Summing it up, developer Rishi writes, a few billion tokens later on data, code, and analysis, here's what I can say. Gemini 4 roughly equals Opus 4.8. Now, going back to anti-gravity, one of the quirks of this release that could be causing the difference in experience is that Google decided to differentiate between pro and ultra subscribers on coding. Pro subscribers only have access to Argon in the Gemini harness. You still have access to the Gemini CLI, but that's not the same as Google's anti-gravity coding harness, which is their equivalent of Codex or Claude Code. To use Argon in anti-gravity, users need to upgrade to the ultra tier. Pavel Hurn again noted a huge difference in performance between the two setups. Running his bug hunt test again, he wrote, Marcelo P. Lima summed up the frustration,

Why does Google make it so hard to try out their new model? I have a workspace account. I can't pay $100 to get AI Ultra add-on because that's only available to individuals. In order to get AI Ultra access, I have to talk to sales? This is crazy. Make it frictionless for people to try your product. If your models have all this intelligence, certainly you can ask Argon to VibeCode a sales page that'll let me buy this add-on. Now, regular listeners of this show will not be surprised that there is a big difference depending on the harness. And frankly, for Google, it would be great news if model performance was a setup issue, not underlying model capability, given how important that agentic coding use case remains. Still, one area that Google's models have always shined is on the multimodal. And given how much every model now has to compete in the arena of social media, where videos and 3D renders capture all the attention, you would think this would be an area for strength for Google. Certainly, the company released some breathtakingly impressive examples alongside the model. Dennis Hassabis showcased a modern version of Battle Chess as an example with the results looking like AAA game graphics and way beyond in many cases what we seen from Rivals Chetisluo was blown away by one of the demos from Google which depicted a massive pagoda and temple complex

from multiple angles, including a full walkthrough. Others noted that this wasn't a one-shot, but a massive multi-agent run that would have been quite expensive, but still the output is impressive. Professor Ethan Malek was able to produce a version of his Imminence game that he's been using to test different models with similarly impressive visual results, but others have found much more lackluster results. Max Techno writes, Guys, I'm genuinely confused about Gemini 4 Argon. Just ran Argon High through my old VibeCheck prompt, and I really don't know how to evaluate it. Now, Max's test is a 3D render of a vehicle that you can look at from different angles. Max continued, It did a really good job with the car's shape. Looks much more like the reference than Opus 5.5's, but the colors look completely washed out, the textures are literally floating, the garage scene is still way too basic, and the UI is still stuck in 2024. It was fast, but be real. Would you actually choose this over an Opus output with all of these cons? Max actually even later wondered whether they trained the model on hyped public tests. He said, it's not close to Opus 5.5 or Astra, and honestly, Soul and Sonnet sometimes look way better.

It's good at rockets, pelicans, and voxel hype tests, but the second it needs to do anything actually creative, it serves junk. So what are we supposed to make of all this? One of the things that I said about Astra when it came out is that it felt like a model that was really asking you to push yourself farther. Not to simply do the same stuff a little bit better, but to really try and do totally different things. And given some of the multimodal and 3D capabilities of Argon, maybe that's a similar way that we should be looking at this model. Certainly the people who were most impressed tended to be exploring these types of dynamic, multimodal, advanced or scientific use cases. Fawfer AI writes, Argon has this amazing natural understanding of the world. It can do so many things from first principles. BionX shared a futuristic 3D battle helicopter made purely in 3JS, saying, Everything you see was procedurally generated from scratch. This is a result of a multi-round self- iteration loop of Argon and Nano Banana. Venture investor Justine Moore wrote, Asked Gemini for Argon for an interactive museum of Apple computers,

and didn't provide any images or data. It scraped archival Mac photos, wrote a tool to clean the images, pulled 50 years of specs, and built a 3D cover flow with UI that morphs to match the decade you're in. In other words, it is a wildly impressive execution of what she was looking for. Web Sim's Surya argued that this is really what the model is about, writing, Gemini 4 Argon is the first model where it felt like an image, video, and language model all at once. And if that's the case, it's not surprising that some of the early tests feel like they're just scratching the surface. The Bug's dev wrote, I feel like we haven't seen enough of what Gemini 4 Argon can actually do. I gave it one image of a Gothic Victorian carriage and asked it to build this whole thing as an interactive 3D experience. Just one image, and this is what you get. A 174,000-face model from scratch, complete with opening doors, hinged windows, working wheel physics, suspension, flickering lanterns, and even custom-sculpted skulls. All of this generated through code, no blender. What's clear is that whatever it's for, Gemini 4 is different. Ethan Mollick writes, Using Gemini 4 makes you realize how many other models must be distilling Claude or GPT

or using a similar training pipeline. It feels quite different than other models. Theo agrees that it's different, but doesn't necessarily love that, writing, It's a huge jump for Gemini and pretty good value for the price, but it's still a Gemini model, though. Does weird stuff regularly, and I don't feel like I can trust it at all. Kavish Krishnakumar tested it extensively with Agenda Coding and wrote, I've been coding with Gemini 4 Argon today, and I actually completely forgot about the character traits Gemini models have when coding. They just love to structure code in a certain way and add comments. Not too long either. Mostly neat one-liner comments, which is a good thing, if anything. I kind of like it. I'm not even certain if the code works yet, which is the classic Gemini trait, but it's just nice to look at code if that makes sense. When you compare it to the other models, their habits are wildly different. It's kind of refreshing to talk to a model that feels unique to its own model lineage. Expanding out, Google DeepMind's director of AGI Economics, Alex Emas, wrote, One thing that's been underemphasized is the importance of having true diversity of output at the frontier. Using a single model, there are unmistakable signatures of each and everything

from writing to taste and code, and this is most obvious in design, e.g. the color palette chosen as background. I've been using Gemini 4 for some time now, and I can say it is completely unique. This is obvious from the visualizations everyone is sharing now, but also in the way that it writes and communicates back and forth. For an economist, the benefits of diversity and independence are baked into nearly all of our models. Correlated signals, information, and priors is how you get fragility, learning traps, and AI slop. So I'm very glad to see our big boy out there for people to use. And bringing it all together, Dan Shipper from Every even found at least one place where it doesn't take some wild imagination to consider switching to Gemini for a core use case. He wrote, Gemini 4 is out and it's surprisingly good? I admit I was ready to trash this model when I saw the benchmarks, but six of us at Every spent the past week testing it across coding, writing, design, and knowledge work, and were pleasantly surprised. It's a breakthrough if your company lives in Google's ecosystem, and a strong option for frontier hoppers as a secondary model for certain tasks.

Some of the things that Dan found was that Gemini 4 was, in his words, shaky at refactors but fast on greenfield apps, basically that it could build new stuff pretty quickly and efficiently and well, but that it, in his words, cut corners on self-verification and existing code bases. He also noted the problems with the harness, calling it a 2026 model trapped in a 2025 And I think it's worth noting that in addition to missing eight months of coding model development, Google has also more or less missed eight months of Harness UX iteration. Dan put it bluntly, saying, But then it comes to writing, where Dan called it his new favorite writing partner. He wrote, DeepMind hired professional writers with red pens to stamp out AI slop, and it worked. I use it to untangle structure and sharpen metaphors. It's got a natural feel for structure and detailed explanations that Astra and Opus lack. The net result, he said,

I keep ChatGPT and Claude open day-to-day and switch to Gemini 4 when I need an editorial partner to structure a tough essay, sharpen a metaphor, or close read a book. Reach forward for live writing collaboration, fast, low-effort prototypes, or steerable UI and dashboard design, ideally via cursor or the API to skip anti-gravity. When a model is good enough to force you to use a bad harness, it tells you something. Given how bad all the other models continue to be on writing, if Dan's results really stick, that alone would be enough for me to spend some time with this model. So, when we come back to the question of where Argon fits, the answer seems to be so far, we're not quite sure yet. That writing use case is enough to give it a test, but it does feel like, as that one ex-poster said, that we may be just scratching the surface of this. And once again, we might want to push ourselves to think about harder, stranger, more unexpected things that we can do with it to really see its full potential. Or alternatively, wait till others do that so I can talk about them on this show and you can just do the stuff that actually works. Either way, I'm excited that we are sitting here on Wednesday in the middle of October

with both Google and Meta meaningfully back in the model race. But for now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always. Until next time, peace. Thank you.