GPT-6.1 Sol in Devin matches GPT-6 Sol on FrontierCode at 81% less cost

Cognition, which builds the coding product Devin, added OpenAI's GPT-6.1 Sol to Devin and reported that it scores about the same as its predecessor on Cognition's FrontierCode 1.1 benchmark for far less money. GPT-6.1 Sol scored 60.4% against 60.7% for GPT-6 Sol, and cost $0.31 per task, which Cognition says is 81% less.

That 81% compares two settings, not only two models. GPT-6.1 Sol's cost is given at medium reasoning effort, a setting that limits how much the model works through a problem before answering, while GPT-6 Sol's is at max effort. The post does not state the effort level behind each score or how the two models compare at the same effort level, so the result shows a cheaper configuration reaching a similar score rather than a like-for-like comparison of the models.

FrontierCode is a benchmark Cognition describes as built with open-source maintainers to measure whether a maintainer would actually merge a model's code change. The figures here are Cognition's results on that test, run inside Devin. The post's attached image and linked page were not available for this explanation, so any further breakdown they give is not covered here.