Opus 5.5 Codes Music Videos: Musk Retweets, Prompt Engineering, Sonnet 5.5 Leak

Opus 5.5 Generates Music Videos by Writing Code

Anthropic's Claude Opus 5.5 has demonstrated the ability to produce full music videos by writing rendering code from scratch. The capability went viral after Elon Musk retweeted multiple examples, commenting "Wow!" "Epic!" and "This time I truly feel AGI." Three standout examples include a five-minute Google biography video with rapid-cut visuals synced to the beat, a two-minute Steve Jobs tribute condensing his life from garage startup to personal computing revolution using 23 transitions at 120 BPM, and a cyberpunk MV created from a blogger's 2011 acoustic guitar track re-arranged into techno via Suno.

The "Donald" Challenge: Same Song, Same Reference, New Visuals

X user donald launched a replication challenge: feed Opus 5.5 the original audio of "Claude Pop" by NotinReality plus the reference MV, then ask it to rebuild the visuals entirely. Donald spoke requirements for five minutes, then left Opus 5.5 to write code, storyboard, generate frames, add transitions, and export a 2:21 video over 12 hours. He subsequently published the ~9,500-character prompt, triggering a wave of derivative works.

Pleometric produced an anime-tech apocalypse theme with pixel-stage UI.

mexicat used brighter, more playful visuals and iterated via Claude for fine-tuning.

anabology added Midjourney assets and moodboards, extending runtime to over five minutes — the version that floored Musk.

Prompt Engineering Is Only 10%; The Production Framework Is 90%

Claude Code team member Thariq noted the prompt alone is 10k characters and includes skills, examples, and API keys. Creator Movez emphasized: "Prompt is only 10%, the other 90% is the production framework." Movez shared concrete techniques that determine output quality:

1. Feed Reference Frames, Not Vague Adjectives

Provide a single frame, a full video, or a portfolio; have the model extract a unified style guide covering palette, typography, lens choices, and transitions so subsequent frames stay consistent.

2. Build Animations as seek(t) Time Functions

Each second maps to a deterministic frame calculation. This enables parallel browser rendering of different segments and allows re-rendering only a broken shot instead of the entire video.

3. Generate a Beat Map Before Visuals

For music-driven MVs, first analyze BPM, downbeats, and volume peaks. Example: "3s — kick hits; 7s — chorus starts; 12s — emotion spikes." All cuts, text animations, color shifts, and SFX lock to these timestamps.

4. Never Render Final Quality on First Pass

Review storyboards and low-res previews, assemble keyframes into a contact sheet, check copy, composition, and motion. Fix the worst three issues, then re-render. Movez's own watercolor short required 163 model calls over nearly seven hours.

Code-first generation is real. One-click finished video? Dream on.

Sonnet 5.5 Benchmark Leak: Pixel Animation Comparison

Anthropic hinted at Sonnet 5.5 and Haiku 5.5 releases within weeks. Leaker JAZII posted a four-model comparison — GPT-6 Astra, GPT-6 Sol, Claude Opus 5.5, and an alleged Sonnet 5.5 — all generating a moonlit cherry-blossom Japanese town in pixel art. Observations from the leaked frames:

Sonnet 5.5 shows denser architectural layering, richer foreground/background separation, and more lavish lighting, blossoms, and micro-details.

Both OpenAI models produce flatter compositions.

Opus 5.5 is more restrained and clean.

This time Sonnet beats both OpenAI models.

If Sonnet 5.5 brings Opus-level aesthetic coherence to a cheaper, more accessible tier, it could significantly lower the barrier for code-driven video creation.

Reference Links

https://x.com/anabology/status/2103534482930491441

https://x.com/_mexicat/status/2103108369569726802

https://x.com/notjazii/status/2103884167104831573

https://x.com/0xMovez/status/2104240559632384173

https://x.com/elonmusk/status/2104360927474921529