Summary
AI Search puts the newly released Claude Opus 5.5 through a broad set of hands-on tests, covering what it can and cannot do along with its specs, pricing and benchmarks. Most tests run through Claude Code with the model set to max effort, which the presenter says avoids spinning up many concurrent agents while keeping the same level of intelligence.
The first round tests real-time adaptation by having the model drive a browser through a gauntlet of CAPTCHA-style puzzles. It handles image-grid selection, rotation, tic-tac-toe, word search and Where’s Waldo, and it works out that whack-a-mole needs a wait-and-click strategy. It struggles with a car-parking task that takes over ten minutes, which the presenter counts as a fail.
Other tests include a from-scratch 3D ray-tracing simulation of a bullet piercing a water balloon, using a separate critic agent to score realism on each iteration, and a reconstruction of an Airbnb listing in Blender through Blender MCP, complete with a virtual-tour camera path. It also compares its answer on Eratosthenes’ calculation of Earth’s circumference with GPT-6 Astra. A sponsored segment covers using Higgsfield MCP so Claude can act as creative director for image and video production.
📺 Source: AI Search · Published September 24, 2026
🏷️ Format: Review
