Summary
Theo of t3.gg goes through how to get the most out of Anthropic’s Opus 5.5, drawing on his own use of the model and a new guide from Addy Osmani, the former Chrome team engineer who recently joined Anthropic. He says the model produces strong code, is pleasant to work with and stays on task through long runs.
A central piece of advice is to hand over the whole task and define what done means. A good prompt states what you want, what completion looks like, and when the model should stop and ask, for example a payments endpoint migration that is finished when every endpoint uses the new client, the old one is deleted and tests pass. He also recommends dropping “think hard” instructions, since the model already decides how much to reason.
The most notable section covers reasoning levels. In his Skatebench runs, moving from X high to max raised average tokens from 338 to about 5,000, average response time from 6 to 50 seconds and the slowest run to 600 seconds, for a gain from 78% to 79% accuracy at roughly 13 times the cost. His conclusion is that lower levels set a ceiling on thinking, while max forces the model to think more, so he keeps X high as his default.
📺 Source: Theo – t3․gg · Published September 25, 2026
🏷️ Format: Deep Dive
