The Catastrophic Math of $200 AI Subscriptions and DeepMind’s 4 Paths to Superintelligence

Yesterday in AI  |  June 16, 2026

The Catastrophic Math of $200 AI Subscriptions and DeepMind’s 4 Paths to Superintelligence

The economic reality of both AI crime and AI commerce is hitting a massive turning point. Today's episode breaks down Google's historic federal lawsuit against "Outsider Enterprise," a Chinese cybercrime ring that weaponized Gemini to launch millions of attacks, exposing the terrifying ease with which AI eliminates the friction of global fraud.
 
Plus, we expose the quietly catastrophic math threatening the flat-rate AI subscription model. New data reveals that full utilization of a $200/month ChatGPT Pro account can cost OpenAI up to $14,000 in raw compute tokens, leaving model providers dependent on users *not* using what they pay for. We also dive into OpenRouter's benchmark-shattering "Fusion" beta, examine Google DeepMind's explicit 60-page project plan to reach machine superintelligence, and break down CISA’s frantic new 3-day emergency patching directive for federal networks.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Tuesday, June 16th, and a Chinese cybercrime network just used AI to scam over 100,000 Americans, forcing Google to file a massive lawsuit with the FBI's backing. We've also got new data showing the economics behind top-tier AI subscriptions are looking quietly catastrophic. A clever workaround that lets you smash frontier model benchmarks, and Google Deepmind publishing what might be the clearest, most explicit roadmap yet to machine superintelligence. Let's get into it. Google filed a sweeping lawsuit on Friday to dismantle a Chinese cybercrime syndicate operating under the name Outsider Enterprise, and the mechanics of the case show exactly what industrial-scale AI-powered fraud looks like. The group weaponized Gemini and other large language models to pump out thousands of hyper-convincing text phishing campaigns, the kind of urgent bank alerts and fake package warnings cluttering your notifications. Behind those texts sat a massive infrastructure, over 9,000 fraudulent websites, a million spoofed URLs, and 2.5 million scam messages blasted to Android devices in just a two-week window this past May. The operation successfully siphoned millions of dollars from hundreds of thousands of victims. What makes this terrifying is how AI completely rewrote the crime syndicate's balance sheet. Historically, generating localized, error-free copyright across thousands of distinct campaigns and spinning up thousands of realistic corporate clones required massive human labor and created operational friction. AI erased that friction overnight. The startup cost for a global scam operation has collapsed to near zero, while the destructive output has skyrocketed. Google's general counsel called this their first coordinated lawsuit of this nature, signaling a pivot toward behind-the-scenes engineering takedowns rather than just playing whack-a-mole with individual accounts. But the real cynical undercurrent here is that defensive tech is bringing a knife to a gunfight. Google rolled out an AI spam detection feature for Android in early June, which successfully flagged 55,000 malicious texts. Meanwhile, this single group managed to deploy 2.5 million messages in half that time. That widening chasm between offensive AI scam capabilities and lagging defensive infrastructure is the most underreported threat in tech today. And while these cyber criminals are extracting millions using basic LLM capabilities, the AI labs building the models are realizing that their own premium subscription numbers don't add up. A devastating report circulating Monday put some deeply uncomfortable math on the table regarding what it actually costs to keep a frontier model running. According to the data, OpenAI's $200 a month ChatGPT Pro Tier could be costing the company upwards of $14,000 per user per month in raw API compute costs if that user hits full utilization. For Anthropics Claude Max, that number hovers around $8,000. The entire flat rate subscription model only survives on the hope that the vast majority of subscribers never actually use the product they're paying for. In fact, OpenAI reportedly hits its break-even point at a mere 11% utilization. If more than one out of every nine pro subscribers decides to genuinely put the model to work, OpenAI loses money on the account. To drive the point home on enterprise scale, one unnamed corporate client managed to burn through $500 million in a single month just running workloads across Claude. Sam Altman has openly admitted that rising token costs are a structural headache, and smart corporate IT teams are already reacting. They are slashing up to 95% of their AI spend by routing basic tasks to cheap open source local models, reserving the brutally expensive frontier models strictly for high-value engineering bottlenecks. The reality is simple: flat rate premium subscriptions are unsustainable. Either prices are about to jump drastically, heavy users are going to get aggressively throttled, or the underlying hardware efficiency has to pull off a miracle. But while the single model subscription economy is looking financially fragile, a clever product update from OpenRouter might have just made single model dependence obsolete anyway. Late last week, OpenRouter launched a new beta product called Fusion, and it is a brilliant piece of commercial timing. Instead of forcing you to guess which elite model is best for your workflow, Fusion automatically routes a single prompt through multiple leading models simultaneously and synthesizes the results. The system maps exactly where the models align, flags where they contradict each other, weeds out individual blind spots, and compiles a single optimized response. While running ensemble methods is a well-known trick in machine learning, productizing it as a consumer-facing tool right when single model subscriptions are bleeding cash is a sharp business move. Open Router claims Fusion significantly leaves individual models in the dust, beating out both GPT 5.5 and Claude Opus 4.8 on a benchmark of 100 complex research tasks earlier this month. If you are an enterprise team looking to optimize your token spend without sacrificing elite performance, this is a beta worth testing. Yet, while developers are busy combining today's models to save a buck, Google DeepMind is looking past the current horizon entirely, charting the actual engineering path to superintelligence. Deepmind quietly published a dense 60-page blueprint titled From AGI to ASI that bypasses the usual philosophical hand waving and outlines four explicit paths to artificial superintelligence. Path 1 is the status quo, pure brute force scaling of compute and data. Path 2 predicts a sharp discontinuous leap via entirely new algorithms that abandon today's transformer architecture altogether. Path 3 is the inflection point, recursive self-improvement, where advanced models are handed the keys to write, test, and optimize their own code and training loops. And path four is collective intelligence, where networks of human-level agents coordinate into a massive hive mind system. The paper also highlights a concept called faithful uncertainty, the idea that an AI must be architected to explicitly hedge its bets when it lacks data, using that uncertainty signal to decide when to stop guessing and start executing a search or calling a secondary tool. DeepMind frames the lack of this mechanism as a utility tax, arguing that today's heavy-handed anti-hallucination guardrails make models too timid to be truly useful. This isn't an academic think piece. It reads like an active internal product roadmap. But as DeepMind maps out the future of autonomous agent networks, federal cybersecurity regulators are realizing that our current defensive timelines are entirely too slow to handle them. CISA dropped a severe new directive last Wednesday that directly mirrors the panic we're seeing in the Google fraud case. Federal agencies now have as little as three days to patch high-risk software vulnerabilities. The new mandate completely tosses out the old, relaxed, calendar-based patching schedules. Now, if a bug is publicly exposed, actively exploited, and easily automated by attackers, the clock hits 72 hours. The reason is simple. AI is being used by adversaries to scan code bases and automate exploit kits at a velocity that traditional bureaucratic IT timelines cannot survive. This lines up perfectly with separate research dropped by NIST last week, which concluded that static safety guardrails are inherently useless against an adaptive AI-armed adversary. If an attacker can iterate infinitely using automated tools, they will find a bypass. For enterprise IT leadership, this marks a massive paradigm shift away from compliance checkboxes and into continuous active red teaming. And while the US government is scrambling to tighten its security timelines, European regulators are taking advantage of the chaos to quietly move their own goalposts. Over in Brussels, the EU Parliament is advancing a digital omnibus legislative package that buys tech companies a significant amount of regulatory breathing room. The update pushes the strict compliance deadlines for high-risk AI systems back to December 2nd, 2027, with the complex safety component mandates delayed all the way out to August 2028. Corporate legal teams across the globe are breathing a sigh of relief, but don't mistake this extension for weakness. The exact same legislative package introduces an immediate ironclad ban on AI-generated non-consensual intimate imagery, explicitly targeting the proliferation of nudifier apps. European regulators are proving that they are moving away from broad, vague risk categories and stepping directly into highly specific enforcement-backed criminal bans. It's a precise regulatory playbook, and if history is any indicator, these specific bands will migrate across the Atlantic to U.S. markets much faster than the broader compliance frameworks. And that's it. If you have any feedback about this show, you can email Mike at yesterdayNai.news, or you can find me on LinkedIn, X, or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious. And I'm taking a week off, so I'll see you in a week.