LLMs: Intelligence vs. Cost

TEXT START: ArtificialAnalysis is a website that benchmarks the intelligence of various LLM models.

The Dissection

This is a technically useful cost-curve analysis with a much larger implication than the author acknowledges. It removes logarithmic distortion, provider pricing noise, and misleading datacenter costs to expose the real frontier: increasingly capable cognition is becoming cheap, replicable, and locally deployable.

The article studies model procurement, not economic survival. Its key claim—that a cheap model is “enough for 90%” of what people need—is not reassuring under the Discontinuity Thesis. It is the employment kill mechanism. If most routine cognitive output can be purchased for cents, employers do not need most routine cognitive workers.

The author sees AI as a cheaper tool. The structural reality is that the tool is becoming a scalable substitute for the people previously required to operate the economy.

The Core Fallacy

The article confuses diminishing marginal value to an individual user with diminishing systemic disruption.

A six-point intelligence gap may be barely noticeable to a casual user, but that is irrelevant to labor displacement. The system does not need the most intelligent model to replace human work. It needs a model that is capable enough, cheap enough, reliable enough, and deployable at scale. GLM-5.3-Flash at $0.023 per task is therefore more economically dangerous than Fable 5.1 at $3.69. The cheap model reaches the substitution threshold for a much larger fraction of work.

The article’s price frontier is evidence for P1, not evidence against collapse. P1 produces P2: institutions cannot preserve stable human-only economic domains when comparable cognitive output is available at radically lower cost. P2 produces P3: the majority lose access to economically necessary labor.

The analysis also treats benchmark cost as the relevant unit. The decisive unit is total cost of producing an acceptable business outcome, including orchestration, verification, tools, latency, liability, and integration. Even where those costs remain substantial, they usually strengthen the case for capital-owned automation rather than restore human bargaining power.

Hidden Assumptions

• Benchmark intelligence scores are treated as reliable proxies for real-world task substitution.
• “90% of what people need” is implicitly treated as economically equivalent to “90% of the labor economy.” It is not. Jobs can collapse when only part of their task bundle is automated.
• Current API prices are treated as meaningful long-term economics despite aggressive competition, falling inference costs, and vertical integration.
• Local hardware is counted as free because the owner already wants it. That is a consumer accounting trick, not a social cost analysis. Capital, depreciation, supply chains, maintenance, and access still exist.
• Cheap open-weight availability is treated as equivalent to dependable deployment, ignoring security, reliability, data, and coordination constraints.
• Intelligence improvements are modeled as smooth diminishing returns. Real workflows can have threshold effects: a model that crosses the line into autonomous project completion is vastly more valuable than one that merely scores slightly higher.
• The analysis assumes the main question is what users can afford, not who owns the compute, models, platforms, data, energy, distribution, and verification layers.
• Human labor remains an implicit fallback option. The entire thesis says that fallback is precisely what loses its scarcity value.

Social Function

Classification: partial truth, transition management, and ideological anesthetic.

It is partial truth because the cost comparisons are materially useful and reveal the accelerating commoditization of cognition. It is transition management because it teaches users how to route around expensive frontier models, exploit third-party pricing, and move suitable workloads onto local hardware.

It becomes ideological anesthesia by translating mass substitution into consumer surplus: cheaper subscriptions, cheaper tasks, and a model that is “good enough.” That framing erases the ownership conflict. Lower prices help buyers; they do not preserve the bargaining position of workers whose output has become reproducible.

This is not pure copium. Its numbers accidentally document the opposite of its social implication. The article is a cost ledger for the demolition of wage scarcity.

The Verdict

The article correctly identifies the most important economic fact: high-grade cognitive capability is becoming cheap enough to be routine infrastructure. It misreads that fact as a consumer optimization story rather than a productive-participation collapse.

The frontier is not merely intelligence versus cost. It is human labor versus replicable machine output, with ownership determining who captures the surplus. The author sees the price curve. They have not followed it to the corpse of the post-WWII employment–wage–consumption circuit.