GPT-5.5 Ships With Medium Effort As Default. The Default Is The Finding, Obviously.

OpenAI, Google, and Anthropic have apparently converged on shared prompting guidance for their 2026 frontier models. The advice is straightforward: define goals, success criteria, constraints, evidence, and desired output rather than micromanaging every step. GPT-5.5 defaults to medium effort, with low effort for simple tasks and high or xhigh for harder agent work.

This validates what competent practitioners have known for years. Over-specification constrains a model's reasoning capacity. The mechanism is delegation: you define what success looks like and let the system figure out how to get there. Think of it like assigning a task to a graduate student. Tell them the destination, not every turn of the route.

OpenAI, Google's Gemini 3 team, and Anthropic independently arrived at overlapping guidance. GPT-5.5 specifically offers effort tiers from low through xhigh for different task demands.