Share with your CTO
Cognition cofounder Russell Kaplan is making a pointed claim: for a growing slice of software engineering tasks, raw model intelligence no longer determines which AI wins. Kaplan calls the phenomenon “intelligence saturation,” and says Cognition’s internal data backs it up. The Information’s reporting adds the quantitative edge: a 10-point model intelligence advantage now drives only 20% more developer traffic, down from roughly 69% in late 2025. When any model clears the competency bar, cost and speed become the entire competitive surface.
What this means for your business
Your AI tooling procurement decisions just changed shape. If your engineering teams are running routine tasks like test generation, boilerplate scaffolding, or code review through an agentic coding platform, the premium you’re paying for the frontier model tier is increasingly hard to justify. The smarter model still matters at the edges, but the edges are shrinking. “Good enough” now covers more ground than most engineering leaders have updated their vendor evaluations to reflect.
The competitive logic for the big labs gets genuinely uncomfortable here. OpenAI, Anthropic, Google, and Meta have all built their developer narratives around capability leadership. Intelligence saturation turns that into a commodity race they didn’t train for. Whoever owns the inference infrastructure with the lowest latency and most aggressive pricing, not necessarily the sharpest model, will capture the volume contracts. That’s a different moat, and most frontier labs aren’t currently winning it.
The signal worth watching: whether this saturation dynamic migrates up the complexity stack. Right now it’s hitting routine tasks. If it reaches architectural reasoning, cross-repo refactoring, or multi-system debugging within 18 months, the entire rationale for paying frontier-model prices in developer tooling collapses. I’d revise this view if frontier models demonstrate sustained, measurable lift on genuinely hard engineering tasks that commodity models repeatedly fail, but the trajectory of The Information’s data points the other direction.
Concept deep-dive: Intelligence saturation
Intelligence saturation describes the point at which multiple competing models all clear the performance threshold required to complete a given task reliably. Below that threshold, smarter wins. Above it, the task becomes a commodity and buyers optimize on price and speed instead. Think of it like broadband in 2010: once connection speeds exceeded what streaming video required, customers stopped caring about raw megabits and started caring about price and reliability. For enterprise AI, saturation means the evaluation criteria for model selection have to be rebuilt around operational economics, not benchmark scores.
Based on reporting from Cognition Cofounder: AI Race Hits ‘Intelligence Saturation’ Point, originally published 2026-10-20 12:56:00.
