Codex vs Claude Code: Sol vs Opus in CoAnimator

We gave Sol in Codex and Opus in Claude Code the same podcast source and kinetic-typography brief. Each chose its own cuts and built a CoAnimator project under 40 seconds around Peter Singer’s comparison of AI to an intern: knowledgeable, inexperienced and eager to please.

Sol’s edit runs 30.9 seconds; Opus’s runs 29.3 seconds. Both retain the original voice while cutting pauses and redundant phrases. Both add layered photographic imagery, selected words, music and sound effects.

The short answer

Compare the moment Singer talks about experience. At 0:11, Sol sets “experience” beside an assembled precision instrument in a dark workshop. At 0:12, Opus places it above loose watch parts and a blueprint. Both cut back to the studio as Singer turns to the human intern.

Watch Sol and Opus edit the same podcast

The shared source is Next Level Podcast, Episode 7, hosted by Anu Venkatesh with Peter Singer. Both edits draw from the passage around 54:44 to 55:43 in the original episode. DVIDS marks the item public domain.

The voice stays at its original speed. Both edits add a restrained music bed and original sound effects. Inter Tight typography and a paper, ink and mint palette connect the compositions. The real studio footage has a warmer, more neutral grade, while the AI-generated metaphor images use separate foreground, object and environment layers.

Sol in Codex

Watch the Sol MP4 · Explore the Sol project player.

Sol uses three generated layers: a workshop, a brass-and-steel instrument and tools close to the lens. The layers move separately to create depth, with the words placed in the empty space to the instrument’s right.

The edit cuts between the instrument scene, Singer’s close-up and the studio two-shot. In one shot, “HUMAN” appears behind the speakers and microphone arms. Music stays under the dialogue, with a few short sound cues at cuts.

Opus in Claude Code

Watch the Opus MP4 · Explore the Opus project player.

Claude Opus 5.5 authored this version, including the dialogue cuts, image prompts, typography and sound. Its four images were generated from Claude’s prompts using Codex’s image generator.

A watchmaking blueprint forms the background, loose metal components occupy the middle plane, and a loupe and tweezers pass close to the lens. Later, the edit cuts to a damaged gear. Focus changes and grain carry through to the studio footage.

Watch the MP4s for the finished edits, or use their project-player links to inspect the animations.

Music in both edits: ‘Signal to Noise’ by Scott Buckley, No Piano Melody mix, used under CC BY 4.0. Each agent selected and mixed its own excerpt with fades and dialogue ducking.

Source: Aaron Lebsack / NIWC Pacific via DVIDS. Edited independently for a CoAnimator demonstration. The appearance of U.S. Department of War (DoW) visual information does not imply or constitute DoW endorsement. See DVIDS reuse conditions.

What changes between the two edits?

The brief that gives both agents a useful direction

The shared brief specifies the source passage, maximum duration, typography, image layers and audio treatment. We supplied the same footage and music track and gave both agents feedback on their edits. Each agent chose its own dialogue cuts, generated image prompts, typography and sound cues. We ran the image prompts and assembled the authored files in CoAnimator.

View the shared podcast-editing prompt

Source and editorial target

Use Next Level Podcast, Episode 7, Aaron Lebsack / NIWC Pacific via DVIDS. Select from the supplied 80-second source window beginning at 54:35 in the episode. Make one coherent edit under 40 seconds, aiming for 30–36 seconds where the argument can land naturally.

Preserve the argument: AI as a knowledgeable but inexperienced intern, eagerness to please, and the request for a logical answer. If retaining desire or human-motive language, retain the explicit qualification that the speaker is humanizing it. Remove filler, pauses and redundant clauses without changing the claim. Do not accelerate, pitch-shift, clone or invent speech.

Map every cut and spoken type event to the source timestamps. Record the source in/out, resulting local in/out and frame indices. Place dialogue cuts on clean word or breath boundaries, retaining natural tails. Keep the real people and studio.

Image direction Request a connected set of three or four photographic layers from the image generator: an environment, a matching transparent object and a foreground detail. Use the actual environment as a reference so light, perspective and material belong together. The agent owns its prompts and composition. The images must clarify the spoken argument, with tactile materials, deliberate lighting and useful negative space. No generated people, replacement podcast studio, text baked into the image, fake statistics or borrowed reference footage.

The reference is a style guide only. Do not reuse its imagery, audio, claims or statistics.

Picture and typography 1920×1080, 30fps, landscape. Use Inter Tight at weights 560 and 700. Graphic palette: paper #F6F6F4, ink #121212, mint #3DFF8A. Photographic colour is allowed. Use strong size contrast, selective mint stress, optical spacing and intentional line breaks.

Compose selected words rather than captioning every sentence. Pair a small spoken setup with a commanding word when it helps. Give faces quiet holds. Keep type clear of eyes, mouths and microphone capsules. Entrances should be quick and intentional, with readable holds and mostly hard-cut exits. No typewriter, karaoke, default glitch, emoji, name straps, progress bars, logos or generic end card.

Composite the generated layers with separate parallax, occlusion and focus. A flat photograph under a zoom does not supply the requested depth. Use visible material texture and restrained grain, with controlled optical motion and deliberate foreground layers. Grade the real footage toward believable warmer skin and neutral whites, keeping the foreground mask and its underlying footage identical in colour. Avoid decorative noise and generic technology backgrounds. A few connected visual ideas are stronger than many unrelated effects.

Use at most one text-behind-person moment, and only with actual source-frame masks. Map the recut local frame back to the original source frame; draw the same decoded picture through its matching alpha mask. Match foreground and background grading, protect the microphones and clear the effect outside its valid source range.

Sound The original voice stays foreground. Add roughly four to eight purposeful, original or verified copyright-clear effects: dry material movement, short low impact or air transition. Tie each cue to an actual visual event. No effect on every word, loud risers or stock robot sound. Add a quiet, licensed instrumental bed, choose an excerpt that supports the argument, and duck it beneath the speaker. Record its source, licence and required credit.

Keep dialogue, music and effects on separate native audio tracks, with sane gains and headroom. Measure the working mix. Do not claim subjective listening without an actual audition.

Editable project Build a deterministic motion-graphics project with one render(t), the __kcControl interface, asynchronous renderFrame and ?t= cold-seek boot. Repeated seeks to the same time must reproduce the same frame. Await decoded footage, fonts, images and the relevant foreground mask. No structural CSS transitions or unseeded randomness.

A frame-accurate raw source-select proxy may carry the cuts and a working dialogue/music/SFX mix for standalone playback. It must contain no finished graphic composition. The native embedded animation is muted; native dialogue, music and SFX tracks play once. Duration, animationDuration, proxy frame count and HTML duration must agree and remain under 40 seconds. The locked animation track stays empty.

Record the source mapping, image prompts, sound provenance and checkpoint targets. Include every cut, key word, foreground-mask boundary and closing frame. Verify the actual local project before calling playback or visual checks complete.

Do not export a finished composed master until asked.

What CoAnimator contributes

The agent authors the animation, while CoAnimator provides the stage, timeline and production tools. The podcast picture lives inside the animation; the edited original dialogue sits on its own audio track, with music and sound effects on separate tracks. That separation lets you revise a visual beat or lower an effect without changing the dialogue edit.

For the next pass, name the change: shorten a word’s hold, move a headline into the empty wall or restore a spoken qualification that was cut. The product-video update guide explains how to revise the affected section.

CoAnimator supports both Codex CLI and Claude Code CLI. The Claude Code connection guide covers setup and a first project; a focused video brief helps turn that connection into a useful creative result.

Real CoAnimator videos: character animation and product films

The following examples show other uses for CoAnimator: Claire Vo’s character loop, Rege’s 3D character template and his YourGPT product film.

Codex: an editable character animation inspired by Rege

Claire Vo’s Winter Feelings project documents an eight-second animation made with Codex, QuiverAI and CoAnimator. She credits Rege’s SVG animation as the inspiration.

Watch the original Codex example and read the workflow.

QuiverAI supplied the SVG illustration; Codex animated it in CoAnimator. The project retains the character’s separate elements, allowing changes to the eyes, head and expression.

Rege’s 3D character animation template

Rege’s original character-animation post shows another direction for an agent-built animation: detailed 3D characters with expressive movement. He describes the template as a starting point for both interactive web animations and videos, with accompanying agent skills.

Watch Rege’s original 3D character animation on X.

A YourGPT product film with a complete story

Rege’s original YourGPT video post follows the product through several customer-support situations. The film cuts from a nighttime setting to customer questions, with narration explaining the product.

Rege identifies Opus 5.5 as the agent he used to make the video in CoAnimator. Watch the original video on X.

For a software launch, combine that motion control with actual product footage. CoAnimator’s demo gallery includes the MCP360 demo and YourGPT Copilot SDK launch film.

Inside the CoAnimator studio

The stage, timeline and agent conversation belong to the same editing workspace. This YourGPT project shows how an animation sits alongside voiceover, music and individual sound-effect tracks, with Claude Code CLI handling project changes in the right panel.

Questions about Sol, Opus and kinetic typography

Choose the edit you can keep improving

When reviewing the two cuts, compare timing and readability as well as individual frames. Does each phrase appear at the right spoken word? Does the type clear the speaker’s face? Can you hear the dialogue clearly over the music?

Give either agent one change at a time: shorten a hold, simplify an image or move a word away from the microphone. Play the surrounding cuts after each revision so a better still frame does not leave an awkward pause or interrupt the speaker.

To try this in CoAnimator, choose a short passage you have permission to use and start with the shared brief above. Keep the source footage and project files so you can revise the edit after the first export.