Cognition president and co-founder Russell Kaplan says the competitive focus in AI coding may be shifting. For a growing number of software engineering tasks, once models can deliver reliable results, further gains in intelligence may matter less than cost and speed—factors likely to shape which tools developers and businesses choose.
In comments posted on October 19, Kaplan said many software engineering tasks had reached what he called “intelligence saturation.” When several models can handle a task well, small differences in capability may no longer determine the choice. Developers and companies are more likely to weigh usage costs, response times and how efficiently work gets delivered.
Recent industry data point in the same direction. A 10-percentage-point lead in model intelligence is now associated with about a 20% advantage in developer traffic. At the end of 2025, an equivalent lead corresponded to a traffic gain of nearly 69%. The figures suggest developers may be becoming less responsive to improvements in model capability alone.
Cognition weighs execution costs for Devin
Cognition develops Devin, an AI engineering tool designed to handle parts of the software development process, including coding and related tasks. Kaplan said the trend was “directionally consistent” with what the company has observed internally, but did not disclose customer, revenue or usage figures.
If multiple models can handle routine software work, competition is no longer just about which system can tackle the most complex problems. The cost and time required for each task, as well as the ability to deploy the tool at scale, also matter. Those measures can affect enterprise budgets and whether developers continue using a product.
Coding remains a selling point for major AI labs
OpenAI, Anthropic, Meta and Google, alongside a growing number of other AI labs, have promoted coding capabilities to attract developers and win enterprise contracts. Differences in code generation, debugging and handling complex engineering work have been key selling points.
As AI agents improve at repetitive and routine software tasks, users may reassess the added value of a more capable model. If a cheaper, faster system can complete the same work, customers may be less willing to pay more or wait longer for a limited improvement in performance.
That does not mean the race to build more capable models is over. Reasoning quality may still matter for complex architecture, extended workflows and software projects where error rates must be low. Kaplan’s comments focused on tasks that several models already handle well; they do not establish that the wider AI coding market will shift uniformly toward lower prices or faster responses. Cognition has not said how it may change Devin’s pricing, model lineup or enterprise services.