Claude Opus 5.5: MICCAI Win & 10-Min 3D History Video in One Hour

Overview

The author, an AI algorithm veteran, shares two major achievements using Claude Opus 5.5 as a daily driver: securing first place on the test leaderboard (excluding organizers' baseline) for the MICCAI 2026 TopBrain medical imaging challenge, and producing a polished 10-minute video explaining the Four Crossings of the Chishui River campaign in roughly one hour via a single Claude Code prompt.

Video Generation Pipeline

The video was created entirely through code orchestrated by Opus 5.5. The model first wrote a narration script divided into 56 segments with unique IDs, synthesized speech per segment using edge-tts to obtain word-level timestamps, then assembled a timeline of 150 subtitles aligned to punctuation and word timestamps, totaling 466.6 seconds of narration.

3D Terrain and Map Rendering

Terrain data came from AWS Terrarium elevation tiles (RGB-encoded elevation). Python scripts stitched tiles, resampled to a lat/long grid, computed hillshade for simulated lighting, applied hypsometric tinting, and produced a texture. The height grid was fed to three.js to generate a triangular mesh with 3.2x vertical exaggeration, making the Chishui River valley and Wujiang Gorge clearly visible.

Dynamic Overlays and Frame Rendering

Red Army routes and enemy force arrows were rendered as 3D lines draped over the terrain; place names and unit labels were HTML elements positioned per frame via camera projection. The page exposed a single renderAt(t) function that draws the scene for a given timestamp. Playwright drove headless Chrome to call renderAt frame-by-frame and capture screenshots; with GPU acceleration, four parallel processes rendered 20,457 frames at ~30 fps in about three minutes. Frames were then encoded to video with ffmpeg.

Procedural Background Music

Background music was also generated programmatically. NumPy synthesized strings, brass, and timpani across ten emotional modes (solemn, march, combat, stealth, etc.), each at its own tempo, with cross-fades between segments. During narration, music volume automatically ducked to 36%.

Fact-Checking via Subagent

Historical facts were retrieved by a dedicated subagent. Each claim was tagged with source and confidence. 103 place names were cross-referenced against OpenStreetMap; 9 uncertain locations (e.g., Qinggangpo, Jiuxikou) were left blank rather than fabricated.

Implications

The author argues that Opus 5.5 dramatically lowers the production barrier for knowledge-intensive video content, threatening creators whose value lies in curating public materials into polished videos. As base model capability grows, such assembly-line and even creative tasks in vertical domains will increasingly be absorbed by the model itself — winning the base model means winning everything.