Alberto Romero cuts through the hype surrounding Anthropic's latest release to deliver a sobering truth: the most powerful AI model in existence is effectively useless for the average person. While headlines celebrate record-breaking benchmarks, Romero argues that for anyone not running multi-million dollar enterprise loops, this "Mythos 5" release represents a widening chasm rather than a universal upgrade.
The Illusion of Universal Access
Romero immediately dismantles the assumption that better models automatically mean better tools for everyone. He notes that while Claude Mythos 5 sits at the top of every major benchmark, "you will not be using it." Instead, standard users are funneled into "Claude Fable 5," a version throttled by safety classifiers. The author frames this not as a security necessity but as a strategic segmentation of intelligence.
The core of Romero's argument is that the utility of these models is entirely dependent on task complexity and budget. He writes, "If you're a developer running loop-based agentic coding tasks across a 50-million-line codebase, the benchmarks are directly relevant to your life and work... On the opposite end of the spectrum, you have people using Claude to draft emails... Fable/Mythos are not for you." This distinction is crucial. It suggests that the "AI revolution" is bifurcating into two distinct economies: one where intelligence scales with capital, and another where it plateaus.
Romero's observation that "AGI is here, just not evenly distributed" serves as a sharp critique of the industry's marketing narrative. He posits that for standard white-collar work, the gains are negligible compared to previous generations. This framing effectively challenges the consumerist hope that a new model release will instantly multiply personal productivity.
A jagged superintelligence is a gift for those living in the peaks and irrelevant for us living in the valleys.
Critics might argue that safety classifiers are essential to prevent harm, regardless of who uses the tool. However, Romero's point stands that the specific degradation of capability for "frontier LLM development" suggests a commercial motive to protect Anthropic's own research advantage as much as public safety.
The Token Economy and Test-Time Compute
The commentary shifts to the economic mechanics driving this divide: test-time compute. Romero explains that the model doesn't get smarter by asking better questions, but by being allowed to "think longer, retry, take notes, iterate." This shift turns time into money in a terrifyingly direct way.
He highlights the staggering costs associated with this new paradigm, noting pricing at "$10/million input tokens and $50/million output tokens," which is double that of previous iterations. Romero cites OpenAI researcher Noam Brown to underscore the danger: "empirically, the plateau [on test-time compute] is very far out." This means there is no natural limit to how much money one can spend before hitting diminishing returns.
This creates what Romero calls a "token rich, token poor" divide. He illustrates this with the rise of "loop engineering," where users design autonomous agents that run indefinitely. The risk is catastrophic for the unprepared; he notes, "one loop that goes for a bit longer than it should can empty your bank account and the bank account of your kids." This is not merely a technical detail but a fundamental change in how value is extracted from AI systems.
The author connects this to the concept of Jevons paradox, where efficiency leads to increased consumption. As models become better at generating code or visuals, users will simply consume more of them. Romero quotes mathematician Terence Tao, who observed that his work style has changed because "it's so easy to generate these things now." The implication is clear: the barrier to entry isn't intelligence anymore; it's liquidity.
Invisible Safeguards and Data Harvesting
Romero also dissects the invisible layers of control embedded in Fable 5. Beyond the obvious safety blocks on cybersecurity or bio-chemistry, he points out a fourth category: "frontier LLM development." This classifier actively degrades responses if it detects a user is trying to improve their own AI models.
He writes that Anthropic estimates this affects only "0.03% of traffic," but the strategic intent is transparent. The model is being weaponized against competitors, ensuring that other labs cannot leverage Mythos 5's reasoning traces to catch up. This reframes the safety narrative from protecting humanity to protecting market share.
Furthermore, Romero exposes a cynical dynamic in the "free two-week" trial period for Pro and Max subscribers. He argues, "You're not getting two free weeks; they're getting two free weeks of you." The conservative tuning of classifiers requires massive amounts of real-world data to calibrate, and this trial window provides exactly that.
Writing—not just the creative kind but plain good writing—is beyond near-AGI models, apparently. I guess dealing with language is too much for language models.
Romero's critique of AI writing capabilities adds a human element to his technical analysis. Despite the hype around reasoning and coding, he notes that "good writing" remains elusive, suggesting that these systems are "either/or intelligences." This limitation highlights that while the models can process vast amounts of data, they still struggle with the nuance of human expression.
Bottom Line
Alberto Romero's analysis is a necessary corrective to the breathless optimism surrounding Anthropic's release, proving that the most significant feature of Mythos 5 is its inaccessibility. The piece's greatest strength lies in reframing "intelligence" as a function of capital expenditure rather than raw capability, exposing a future where AI utility is strictly gated by one's ability to burn tokens. However, the argument may slightly understate the long-term potential for cost reductions that could eventually democratize access to these high-compute loops.