GPT-5.6 used a prompt to close a 30-year gap in convex optimization
Summary
The post compares GPT-5.6 and Claude Fable 5 on an unpublished NP-hard optimization problem, evaluating the impact of a native /goal mode. It finds that Fable 5 performs exceptionally well, while /goal can help in some runs but is not universally beneficial, illustrating that goal-driven prompting can both improve and degrade performance depending on the task and model.