The AI coding wars might be hitting a ceiling. Russell Kaplan, president and cofounder of Cognition, the company behind the Devin software engineering platform, posted that many software engineering tasks are now intelligence saturated.
His unpacking of the term is straightforward. Once any model can do a task well, the relevant differentiators become cost and speed, not raw intelligence. The smartest model loses its edge if a cheaper, faster competitor clears the same bar.
The Signal Behind the Claim
Kaplan was responding to reporting from The Information. Their data point: a 10-point intelligence advantage now drives just 20% more developer traffic, down from nearly 69% in late 2025.
He called that finding directionally consistent with Cognition’s internal data. If the trend holds, it shifts the strategic calculus for OpenAI, Anthropic, Meta, Google, and the rest of the AI labs fighting for developer mindshare and enterprise contracts.

What This Means for Your Stack
If you’re picking models for routine coding tasks, the implication is practical. A slightly less capable model that costs a fraction of the frontier tier and runs faster might be the better default. Reserve the top-tier API calls for tasks that genuinely require the extra intelligence.
The saturation point will not be uniform across all tasks. Complex architecture decisions, multi-file refactors, and novel problem solving still favor the smartest models. Boilerplate generation, test writing, and docstring updates might already be in the saturated zone.
Watch your own data. If you cannot measure a quality difference between model tiers for specific task categories, you are probably overpaying for intelligence you do not need.
