How Prompt Tuning Improved GPT-5.5 in VS Code (code.visualstudio.com)

In a recent experiment, the VS Code team collaborated with OpenAI to enhance the performance of GPT-5.5 through "prompt tuning" in a coding context. The focus was on making the AI agent more efficient by tweaking how it explores code before making edits. By modifying system prompts, the team aimed to improve the model's token usage, which translates to lower costs and faster response times, especially important with usage-based billing in place. The experiment tested two new prompt structures against a control version, revealing that a more structured approach (Treatment B) significantly improved latency and reduced overall token consumption. The results highlighted Treatment B, which emphasized a structured approach to coding tasks that encouraged less exploratory behavior and faster validation of edits. This variant led to a 5.68% reduction in average time to the first edit and a significant decrease in the number of tokens used for high-load requests. Although there was a slight decrease in code retention at first, the overall efficiency improvements led to the decision to adopt Treatment B as the new default system prompt for GPT-5.5 in VS Code. This advancement illustrates the importance of prompt design in optimizing AI behaviors, providing a more efficient and cost-effective experience for users.