Physix Frontier · News Briefing Card (Hacker News · Oct 13, 2026)

Cognition Launches FrontierCode 1.1, Top Models Fail Over 40%

KEY FACTS

Cognition released FrontierCode 1.1, which evaluates whether AI edits to real open-source repositories would be merged by maintainers.
The benchmark contains 150 tasks, with a main set of 100 tasks, each run three times.
Claude Opus 5.5 averaged 52% and a 58% pass rate, the highest currently.
GPT-6.1 Sol costs $0.81 per task and scores 49%, the best value.
Of the 100 tasks, 9 were passed by no model, and no task was passed by all models.

KEY DATA

52%Highest average score

58%Highest pass rate

$0.09Lowest cost per task

$12.00Highest cost per task

PHYSIX OBSERVATION

FrontierCode raises the evaluation standard from "tests pass" to "maintainers are willing to merge," getting closer to real engineering needs. Top models still fail more than 40% of the time, showing AI-written code is still far from directly usable. Costs differ by a hundredfold, so when choosing a model, cost-effectiveness deserves more attention than absolute scores.

Source: Hacker News report