Opus 5.5 Music Video Test · Low to Max and Ultracode

Quick Summary

• Devarf ran the same 79-second excerpt of Mirror Mirror by Anno Domini Beats through Claude Code with Opus 5.5 at Low, Medium, High, xhigh and Max effort, then performed a sixth run with Ultracode. The brief asked for an official-feeling music video made entirely from 2D or 3D motion graphics, with dance or story and no filmed footage.
• The controlled variables were unusually strong for a creator experiment. Each run had its own folder, the same song analysis, the same rendering tools, the same open prompt and instructions to work autonomously. The principal intentional change was effort, although Ultracode introduced a different orchestration system and was accidentally paired with Max effort.
• Low produced a complete 2D video in about 42 minutes, used roughly 78,000 tokens and carried an estimated API-equivalent cost of about US$3. Max produced a 3D video in almost three hours, used roughly 625,000 tokens and carried an estimated API-equivalent cost of about US$49.
• Medium through Max all chose 3D. Devarf saw progressively stronger modeling and movement across those runs, but he did not declare one universal winner. Low remained a credible result because its 2D direction suited the task and cost far less.
• Ultracode is not simply another effort level. It adds dynamic workflow orchestration and multiple agents. In this test it mixed a 2D character with a 3D stage made from playing cards, used at least 2.5 million tokens and reached at least US$348 in estimated API-equivalent cost before Devarf stopped the unfinished correction loop.
• Claude did not directly hear the raw song in this setup. Devarf first ran a custom audio-analysis tool that described beats, notes, instruments, dynamics, lyrics and narrative structure. That shared analysis became the timing and story bridge for every run.
• The practical lesson is not that Max always wins. A precise art direction, a stable audio map, reusable rendering tools and an explicit review loop can matter more than raising effort on a vague brief. Higher effort should be reserved for shots whose complexity or failure cost justifies it.
• The video contains bright, rapid flashes. Creators should test the finished animation against accessibility thresholds and modify unsafe sequences. A warning is useful context, but it is not a substitute for reducing or removing hazardous flashing.

Credit, scope and photosensitivity warning

The embedded video was created by Devarf and published on October 5, 2026. He designed and ran the experiment, supplied the tool folders and audio analysis, reviewed all six outputs and reported the time, token and estimated pricing figures discussed here.

Some examples contain bright, fast flashes. The W3C guidance for prerecorded video says content should avoid more than three flashes in any one-second period unless the flashing remains below the defined general and red-flash thresholds. Test and revise the actual frames. Do not rely on a warning alone.

This article is a structured reading of a creator experiment, not a controlled scientific benchmark. The six runs share a useful core setup, but stochastic generation, local hardware, rendering failures and Ultracode's multi-agent behavior introduce variation that a larger repeated trial would need to measure.

The controlled experiment: one song, one brief, six runs

Devarf used a 79-second excerpt of Mirror Mirror by Anno Domini Beats. Each run started in its own folder and received the same source information, audio analysis, rendering utilities and open-ended creative prompt. He ran Claude Code 2.1.287 in the terminal with Opus 5.5 between October 2 and October 4, 2026.

The first five conditions were Low, Medium, High, xhigh and Max. The sixth was Ultracode. All six returned a full-length music video with the original audio aligned to the final render. The experiment deliberately avoided filmed footage, so every visual had to be generated as programmatic motion graphics.

Devarf also told the system to work autonomously from start to finish rather than pause for questions. That makes the test useful for observing initiative, but less representative of a normal directed production in which a person approves the concept, storyboard and expensive revisions.

How the agent understood a song it could not hear

In this workflow Claude Code did not listen to the raw audio file. Devarf used a custom analysis tool, itself created with Opus 5.5 High, to convert the song into structured information about beats, notes, instruments, dynamics, lyrics and story. Every run received that same analysis.

This is the hidden control layer. A useful music-video brief needs a time map: section boundaries, downbeats, lyric entrances, intensity changes, pauses and narrative turns. Without it, the animation may be technically synchronized to duration while missing the emotional rhythm of the track.

The custom analyzer reportedly used about 5 GB of memory on a machine with 6 GB of VRAM. That figure describes Devarf's implementation, not a universal requirement. A lighter workflow can use a beat detector, a transcription model and a compact JSON cue sheet, then let the animation agent work from those outputs.

The shared motion-graphics toolbox

Every folder included tools for Three.js, Canvas 2D, SVG and frame rendering. Those options let Claude choose a visual language while retaining deterministic control over timing, geometry, typography and every rendered frame.

The experiment did not include Devarf's own creative guides or example videos. This matters because the result came from the shared technical environment and the loose prompt rather than a tightly specified brand system. It also helps explain why different runs selected radically different visual strategies.

Programmatic motion design separates construction from encoding. The agent writes the scene logic, renders the sequence frame by frame and assembles it with the soundtrack. A reusable framework can reduce setup cost, while a contact-sheet loop lets the agent inspect representative frames before spending time on another full render.

What changed from Low to Ultracode

Low chose a 2D treatment. It is the fastest and cheapest measured endpoint in the video, yet Devarf still considers it a good result. Its value is a reminder that a coherent art direction can beat raw technical ambition when the brief does not demand depth, simulated cameras or complex 3D assets.

Medium, High, xhigh and Max all chose 3D. Devarf reports a visible progression in model quality and movement toward Max. The transcript does not publish exact intermediate time, token or dollar figures, so this article does not manufacture them. Watch the embedded comparison to judge composition, choreography, rhythm and taste for yourself.

Ultracode combined a 2D character with a 3D playing-card stage. It produced an inventive output, but the process became operationally inefficient. Multiple agents built their own caches, work restarted after usage limits and the system continued trying to correct the project even after it had delivered a playable video.

Time, tokens and estimated API price

Low took about 42 minutes, consumed roughly 78,000 tokens and was estimated at about US$3 in API-equivalent usage. Max took almost three hours, consumed roughly 625,000 tokens and was estimated at about US$49. From these two reported endpoints, Max used about eight times as many tokens and took more than four times as long.

Ultracode used at least 2.5 million tokens and reached at least US$348 in Devarf's estimate. It consumed nearly three times his five-hour usage allowance and still did not reach a clean terminal state. He stopped it after receiving the output because the agents continued entering correction cycles.

Token totals do not reproduce a bill on their own. Anthropic lists separate Opus 5.5 prices for input, output, cache writes and cache reads. The same total-token number can therefore produce different costs depending on the mix. Treat the video's dollar figures as reported estimates, not a transferable quote for your machine or account.

What the official effort settings mean

Anthropic's current guidance presents Low and Medium as faster choices for routine work, High as the balanced setting, xhigh as deeper reasoning for longer coding and agentic tasks, and Max as the most thorough option with the greatest usage. In Claude Code, the setting can be changed with the effort command.

Opus 5.5 uses thinking by default, so effort changes the reasoning budget rather than switching reasoning on or off. A higher setting can spend more time exploring, debugging and refining. It does not guarantee better taste, a more suitable style or a proportionate improvement in the final frames.

A practical default is Medium or High for the first complete pass. Move selected failures to xhigh or Max only when the rubric identifies a problem that more planning is likely to solve. If the issue is vague direction, a missing reference or a broken renderer, increasing effort may only make the wrong approach more expensive.

Ultracode is orchestration, not the next rung on the ladder

Ultracode is a Claude Code mode backed by dynamic workflows. It can coordinate multiple agents and adapt the workflow around a task. That makes it categorically different from Low through Max, which are effort settings for model reasoning.

Devarf accidentally combined Ultracode with Max. He expected the mode to use xhigh by default, so the run became an especially aggressive test rather than a clean sixth effort condition. The multiplying caches, restarts and independent attempts explain why the usage curve jumped far beyond Max.

His recommendation is not to use Ultracode for this music-video workflow. If you still want to test it, begin with a disposable short excerpt, low effort, one active deliverable, a small agent cap and explicit stop conditions. Do not point an unrestricted multi-agent run at an open-ended quality goal such as keep improving until perfect.

The real comparison is direction versus compute

Devarf cannot choose one absolute winner. He sees stronger 3D construction and movement as effort rises, but the Low result remains creatively viable. The experiment therefore supports a curve rather than a crown: extra compute can buy more exploration and repair, while style fit and concept clarity remain separate decisions.

The original prompt intentionally leaves character, palette, camera, typography, narrative beats and quality criteria unspecified. That openness exposes the model's initiative, but it also makes each run solve a different creative problem. A commercial benchmark should lock the storyboard and visual system before comparing execution quality.

For most production teams, spend the first budget on references and acceptance criteria. Define the hero idea, visual grammar, beat map, safe margins, typography, palette and ending. Then compare effort on the same approved storyboard. That reveals whether the model is executing better instead of merely proposing a more expensive concept.

A safer reproducible benchmark

Choose a 15 to 20 second excerpt before attempting 79 seconds. Build one shared audio cue sheet and one approved storyboard. Copy the same project template into separate run folders. Pin the model, Claude Code version, dependencies, resolution, frame rate and render command.

Set a rubric before generation: beat synchronization, lyric or story alignment, composition, legibility, movement quality, continuity, technical correctness and accessibility. Score blinded exports rather than folder names so reviewers do not know which effort level produced each video.

Cap every run by wall-clock time, token budget and revision count. Require a playable draft before refinement, then permit the same number of repair passes. Record agent time separately from rendering time and human cleanup. Repeat the test at least three times before claiming one level reliably wins.

Flash safety must be part of motion QA

Music videos often use rapid cuts, full-frame white flashes and high-contrast pulses on the beat. These can create a seizure risk. W3C's accessible video guidance uses three flashes per second as a central boundary, with additional thresholds for flash area, luminance and saturated red.

Export a low-resolution review copy and run an automated flash analysis where possible. Inspect the actual cut, not just the storyboard. Replace unsafe full-screen flashes with camera motion, color shifts, scale accents, localized highlights or slower transitions that preserve the musical emphasis.

Keep a visible content warning when fast or intense imagery remains, but do not treat the warning as compliance. The safer delivery is the revised video that passes the threshold, accompanied by a version log showing what was tested and changed.

Music credit and licensing checks

The track in the experiment is Mirror Mirror by Anno Domini Beats from the YouTube Audio Library. YouTube explains that some Audio Library tracks use a Creative Commons license and require attribution, while others use the standard Audio Library license and do not. Check the license shown beside the exact download in your account.

Audio Library music is intended for YouTube use under YouTube's terms. If the finished video will also appear in client work, advertising, broadcast, an app or another social platform, review the current license and distribution rights before publication. Preserve the track title, creator, download date and license text with the project.

For a benchmark you can also use music you created or commissioned, or a track with an explicit cross-platform commercial license. This removes uncertainty when the test later becomes a portfolio piece or paid deliverable.

Verdict · use effort selectively and direct the work

Low proved that a complete, coherent 2D music video does not require the largest reasoning budget. Medium through Max showed the value of more reasoning for ambitious 3D construction and movement. Ultracode showed that orchestration can amplify both creativity and waste.

Start with Medium for exploration or High for a demanding first pass. Use a clear audio map and storyboard, review a short proof, then raise effort only for specific problems. Reserve Max for shots that justify the additional time. Keep Ultracode behind strict caps until its multi-agent behavior is part of what you actually need to evaluate.

The useful takeaway is creative leverage, not an effort leaderboard. Claude Code can turn an audio brief and motion toolkit into a full-length rendered piece, but the most reliable gains still come from human art direction, bounded iteration and evidence-based review.