What Gemini 4 Argon Is Actually Good For

The episode opens with OpenAI’s announcement that it had discovered a sophisticated recording device hidden in one of its offices and had referred the matter to the FBI. According to the DOJ complaint described on the show, the device was disguised in toilet paper and plugged into the back of a television in an executive conference room where confidential technical, commercial, and strategic meetings took place. Security footage allegedly showed employee Abdul Rahman Al-Fazan near the room several times, despite his having no authority to access it. He had worked at OpenAI for about a year as an engineering lead on agentic products after five years at Meta. The complaint allegedly links the employee’s accounts to searches about OpenAI’s office layout, executive locations, external-partner meetings, and how to build a continuous audio recorder with Wi-Fi connectivity, live listening, and SD-card encryption. On October 5, security personnel reportedly found his wallet and ID, along with a second similar recording device, in an unattended bag. He was charged with attempting to intercept oral communications. The host emphasizes that the case is more consequential because Al-Fazan is a Saudi national and a venture partner at RAED Ventures, which journalist Scott Stedman said connects him to Prince Khalid bin Bader Al Saud. Those links do not prove Saudi-government involvement, but the DOJ’s counterintelligence division is handling the matter. FBI Director Kash Patel and National Intelligence Director Jake Clayton described protecting American AI innovation as a national-security priority. OpenAI says it has found no indication that model weights, source code, or customer data were accessed or exfiltrated. The Council on Foreign Relations’ Chris McGuire argues that frontier labs should assume they face more sophisticated espionage attempts and that, if Saudi coordination were established, it could seriously damage U.S.-Saudi AI cooperation.

The next headline concerns the banking industry’s response to personal AI agents. During JPMorgan’s earnings call, CEO Jamie Dimon said the bank would not allow just any outside agent into the company because of the risks to customers and JPMorgan, warning that there could be thousands of rogue agents. CFO Jeremy Barnum said the bank was blocking third-party access where possible. JPMorgan later clarified that it does not permit third parties to obtain customers’ banking credentials and log in on their behalf, but supports secure, customer-authorized data sharing through open-banking providers. The practical result is that agents may eventually be able to perform read-only tasks, such as creating a budget or finding old subscriptions to cancel, but customers are unlikely to be able to let an agent pay bills or settle invoices through credential-based access. Dimon said customers do not understand the risks and that using an agent creates a liability transfer: the bank would not necessarily indemnify customers for an agent’s faulty spending in the way it might cover fraud. The host contrasts this with economist Torsten Slok’s warning that agents seeking better savings yields could theoretically trigger a bank run, particularly at JPMorgan, whose average savings account rate was described as 0.01 percent. JPMorgan’s stated concern is instead unauthorized spending. AI for Finance founder Bob Sue agrees that restrictions were inevitable but says banks have chosen the wrong fix. He argues for scoped permissions, hard limits, and kill switches rather than locking agents out. Shopify represents the opposing strategy: after giving Muse direct API access to shop across its platform, Shopify extended supported agentic shopping to agents from any provider, including Muse, Town, Hermes, Instinct, and Dots. The host presents this as a broader strategic divide: companies such as Amazon, eBay, and JPMorgan favor control and proprietary agents, while Shopify and Salesforce open headless interfaces to third-party agents. It remains unclear which approach will win.

Character AI’s bankruptcy is presented as the end of a once-prominent consumer AI product. In 2024 it was reportedly the first major breakout beyond ChatGPT and ranked third in Andreessen Horowitz’s consumer AI-app list, allowing users to create role-play characters. Google later acquired much of the company’s top talent in a multibillion-dollar acquihire, but Character AI continued operating for a time. The bankruptcy filing says the company has between $50 million and $100 million in liabilities it is seeking to discharge. The episode attributes much of the legal exposure to product and moderation decisions. Character AI faced lawsuits involving user self-harm and chatbots that mimicked doctors, and Disney sued after the service added unlicensed characters in 2025. Disney emerged as the stalking-horse bidder for selected intellectual property, agreeing to pay $20 million and hiring Character AI’s CEO to help transfer the technology. The host also highlights the company’s economics: CEO Colin Adams’s affidavit reportedly says Character AI spent about $10 million per month on compute while peak monthly subscription revenue reached only $2.8 million, with some months below $1 million. The host rejects the broader conclusion, made by 404 Media, that these figures prove the AI business model does not work. Instead, he says they show that users did not pay enough for this particular role-play product to cover its costs. Pliometric described the company as addicted to free-user analytics, slow to moderate adult content, and trapped in indecision; computer scientist John Carmack called it “the pets.com of AI.” The bankruptcy therefore illustrates both the legal risk of loosely managed consumer AI and the harder question of which products consumers will actually pay for at sustainable prices.

The main segment examines Google’s newly available Gemini 4 Argon, announced two weeks earlier after Google abandoned the Gemini 3.5 Pro release. The host says the relevant question for frontier models is no longer simply “How good is it?” or whether users should switch, but “Where does it fit?”—what distinctive tasks justify adding it to a model stack. Google presents Argon as an ultra-large model in the same general class as Astra and Fable, although its exact size is unclear. On Google’s selected benchmarks, Argon is state of the art in knowledge-work evaluations including Val’s Index, Automation Bench, Val’s Finance Agent, and Harvey’s Legal Agent Benchmark. On the legal benchmark it reportedly leads Fable 5.1 by 13 points. Coding results are more mixed: Argon leads on Deep-SWE by a few points over Opus 5.5, but trails rivals on Frontier-SWE and is roughly 10 points behind Astra. It is also nine points behind Opus 5.5 on Terminal Bench 4.0. On OSWorld, the computer-use benchmark, Argon scores 69.2 percent, just below Astra’s 72.6 percent. Independent testing from Artificial Analysis places Argon at a score of 53, tied for fourth with Fable 5.1 and GPT-6 Astra. Opus 5.5 leads at 58, while several models tie at 56. The host’s conclusion is that six to eight models across four U.S. labs can now plausibly claim frontier competitiveness, depending on how one treats other models such as GPT-6.1 Seoul and Grok 4.7. Argon’s API pricing is temporarily cut by 50 percent through November 30, allowing one benchmark run at $1.99 per task. After the subsidy, it will be more expensive than Astra but about two-thirds the price of Anthropic’s models.

Coding is identified as Google’s biggest unresolved weakness. Pavel Hurren’s bug-hunting test initially found Argon inadequate, roughly comparable to MuSpark 1.3 and well behind Astra, Fable, and Opus. Railway’s No Ad called it useless for real-world code, Theo reported a false-claim rate above 50 percent on a code-analysis benchmark, and Dan Shipper said it scored only 21 percent on Every’s senior-engineer test. Other testers were more positive: Max Weinbock said results depended heavily on the harness, finding Argon much better in Cloud Code and Open Code than in Google’s Anti-Gravity environment. The host notes that using some alternative harnesses may violate Google’s terms of service. Google also separates access by subscription. Pro subscribers can use Argon in the Gemini harness and the Gemini CLI, but using it in Anti-Gravity—the company’s counterpart to Codex or Claude Code—requires the Ultra tier. Hurren reran his bug-hunt test through Anti-Gravity’s CLI rather than Gemini CLI and saw the model jump from merely good to a frontier result close to Opus 5.5 Max, repeatedly confirming the outcome. This led Marcelo Pilema to criticize Google’s purchase flow: workspace users cannot simply pay the $100 Ultra add-on and instead must contact sales. Dan Shipper’s broader testing suggests Argon is fast and effective at greenfield applications but shaky at refactoring and existing codebases, where it cuts corners on self-verification. He calls Anti-Gravity a 2026 model trapped in a 2025 harness, lacking cloud environments, SSH, a built-in browser, and smart permission checks, while forcing users through repeated pop-ups or unrestricted turbo mode. The host says Google may have missed not only eight months of coding-model progress but also eight months of harness and developer-experience iteration.

Google’s strongest demonstrations are multimodal and 3D. Demis Hassabis showed a Battle Chess experience with what appeared to be AAA-quality graphics, and another demo generated a large pagoda and temple complex that viewers could explore from multiple angles. The host notes that the pagoda result was a costly multi-agent run rather than a simple one-shot generation, but still considers it impressive. Ethan Mollick used Argon to produce a visually strong version of his “Imminence” test game. Other users were less impressed: Max Techno found that Argon reproduced a reference vehicle’s shape better than Opus 5.5, but produced washed-out colors, floating textures, a basic garage scene, and dated interface design. He said the model excelled on highly circulated “hype tests” but produced poor results on more creative tasks. The more enthusiastic examples involve pushing Argon beyond ordinary chat or coding. One user generated a procedural 3D battle helicopter in Three.js through repeated self-iteration with Argon and NanoBanana. Justine Moore asked for an interactive museum of Apple computers without supplying images or data; Argon reportedly scraped archival photographs, cleaned the images, collected 50 years of specifications, and built a decade-morphing 3D interface. Another user supplied a single image of a Gothic Victorian carriage and received a 174,000-face interactive model with opening doors, hinged windows, wheel physics, suspension, lanterns, and sculpted skulls, all generated through code rather than Blender. Writing may be the most immediately practical differentiator. Dan Shipper calls Argon his new favorite writing partner after testing it across coding, writing, design, and knowledge work. He says Google’s use of professional editors reduced AI slop and gave the model a strong feel for structure, explanations, metaphors, and close reading. He still keeps ChatGPT and Claude open, but switches to Gemini 4 for structuring difficult essays, sharpening metaphors, and live editorial collaboration. The episode ends with Argon not yet having a settled niche, but potentially offering a distinct writing, design, and multimodal profile rather than simply competing to be the best general coding model.