diff --git a/README.md b/README.md
index 772cfe4..42f98a5 100644
--- a/README.md
+++ b/README.md
@@ -33,9 +33,7 @@ eval (see benchmarks).*
 
 ## Demo
 
-[![pxpipe vs plain Claude — side-by-side A/B on Opus 4.8](https://raw.githubusercontent.com/teamchong/pxpipe/main/docs/assets/pxpipe-ab-test-thumb.jpg)](https://github.com/teamchong/pxpipe/blob/main/docs/assets/pxpipe-ab-test.mp4)
-
-*Side-by-side — plain Claude (left) vs pxpipe (right), both on **Opus 4.8** (opt-in; pxpipe is tuned for Fable, currently offline). Click to play.*
+▶ **[Watch the side-by-side demo](https://drive.google.com/file/d/1yZYEjr9765aswomG8mopmp3GlRdSLWM1/view?usp=sharing)** — plain Claude (left) vs pxpipe (right), both on **Opus 4.8** (opt-in; pxpipe is tuned for Fable, currently offline).
 
 - **Demo 1 — fix a failing test suite:** both pass; the dashboard shows pxpipe cut the request to a fraction of the tokens (real, server-measured **context/token reduction**).
 - **Demo 2 — a big file-context (40 files, ~382k tokens) plus a math question and a "count this phrase" task:** the math answer (a small **text** needle) reads on both. The phrase-count needs reading the **imaged** filler — so pxpipe-on-Opus can't read it and **honestly surfaces that it won't fabricate a number** (the documented lossy limit: exact values stay text). Plain, meanwhile, bogs down counting file-by-file.
diff --git a/docs/assets/pxpipe-ab-test-thumb.jpg b/docs/assets/pxpipe-ab-test-thumb.jpg
deleted file mode 100644
index bd690f9..0000000
Binary files a/docs/assets/pxpipe-ab-test-thumb.jpg and /dev/null differ
diff --git a/docs/assets/pxpipe-ab-test.mp4 b/docs/assets/pxpipe-ab-test.mp4
deleted file mode 100644
index cf83ae9..0000000
Binary files a/docs/assets/pxpipe-ab-test.mp4 and /dev/null differ