AI Coding Enters "Intelligence Saturation" Phase: Cost and Speed Replace Capability as Core Competitive Factors
nashnova research
Cognition co-founder Russell Kaplan says a growing share of coding tasks have reached 'intelligence saturation' — any major model can do them, so the race shifts from who's smarter to who's cheaper and faster, pressuring the pricing and strategy playbooks of OpenAI, Anthropic, and their rivals.
What does 'intelligence saturation' mean?
Russell Kaplan, co-founder and president of Cognition — the startup behind AI coding platform Devin — said Monday that more and more software-engineering tasks have hit "intelligence saturation."
In plain terms = once any mainstream model can handle a task, "who's smarter" stops mattering. The only differentiators left are cost and speed.
This means → the core metric in AI coding is flipping from "capability ceiling" to "operational-efficiency floor."
What data backs this up?
A figure reported by *The Information*: for every 10-point gain in model intelligence, the resulting developer-traffic boost has dropped from ~69% in late 2025 to roughly 20% now.
This means → the same jump in smarts attracts nearly two-thirds less attention than it used to — diminishing returns on intelligence are setting in fast.
Kaplan said the trend is "directionally consistent" with what Cognition sees internally.
What does this mean for the big AI labs?
OpenAI, Anthropic, Meta, and Google have all pitched coding ability as a headline selling point to win developers and enterprise contracts.
But as models converge on routine tasks, a marginal intelligence edge in one model may no longer offset another model's advantage in price and response speed.
This reflects a deeper signal: the intelligence premium is fading, and competition is shifting toward operational efficiency and pricing strategy — more cost war than tech war.
市场有风险，内容仅供研究参考，不构成投资建议。