[Observation] Where Is Opus 5.5 Strong in Making Video? Notes from a Small Studio

1. Opening

On September 22, 2026, Anthropic released Claude Opus 5.5. Very quickly, posts praising its video-making ability flooded X and Reddit. And yes, a lot of the work people shared was genuinely impressive.

The way it makes video is not complicated: the model writes web animation code, a browser captures the frames one by one, and then a video tool stitches them into a finished clip. You can add sound effects, voiceover, and music along the way.

Models before Opus 5.5 could do this too. So why was it Opus 5.5 that triggered all the excitement? Is "Opus 5.5 is great at making video" just hype?

We went into this exploration with three goals. First, we wanted to know where the ceiling is for models making video through code. Second, our own studio needs a faster video workflow. Third, we wanted to find out whether Opus 5.5 really is that good.

Here is our conclusion up front: Opus 5.5's edge is real. It is genuinely better than other models at making coded video. Details below.

2. How We Did It

When the Opus 5.5 video trend first took off, most posts were just a single prompt. There was no detailed pipeline.

Based on our own video production experience, we sketched out a workflow for making animated video with Opus 5.5, then tested and refined it with short clips. Within two or three hours we had a pretty stable output.

The workflow has ten steps, but from an engineering perspective it is really just two layers.

Layer one: planning and design. Pick the topic (we choose the topic ourselves), write the script, then design the visuals so it actually looks good. The visual plan is not something we came up with from scratch — it draws on best practices collected from many great teams. In the AI era, gathering those best practices is no longer hard. The plan is organized shot by shot, tied closely to the script, and includes sound design and animation design. We call this whole package the "blueprint" internally.

Layer two: coding and rendering. The model follows the blueprint, writes the code, and renders the video.

For the three episodes we have finished so far, Opus 5.5 did layer one and OpenAI's GPT 6.1 did layer two.

Episode two, Synchronization, full video: planned and designed by Opus 5.5, coded by GPT 6.1.

At this point, the workflow was mature enough for our studio's own production. But one question was still open: is Opus 5.5 really that outstanding? If so, which part of its ability makes the difference, and which steps can be handed off to other models?

3. Swapping Out the Designer

When we started episode three, Phyllotaxis, we first wrote a topic brief: what subject this episode covers, which historical events to include, which to leave out, and what conclusion to end on. Think of it roughly as an outline for an article.

Since the ten-step workflow was already built and iterated on, in theory we could hand it to any capable model. We first tried GPT 6 Astra, but after testing a few short shots we found it used up our quota too fast, so we switched to GPT 6.1 sol.

GPT 6.1 ran through all ten steps and rendered a complete video. But the result was not good: the visuals stayed stuck on flat dot patterns; the subtitles lined up with the visuals only mechanically. Viewers could not tell what was going on, and there was no sense of beauty.

Same workflow, but not every model produces a good video.

So we changed the division of labor: let Opus 5.5 finish layer one and output the blueprint; generate a background music track on Suno; then hand the blueprint to GPT 6.1 for coding and rendering. The result was good, and GPT 6.1 used very little quota.

The difference is most obvious if you put the openings side by side. Both start from a blank screen and lead into a sunflower head. GPT 6.1's design starts by placing dots one by one to form the flower head. Opus 5.5's design first grows a young shoot with leaves in the light, the camera rises from eye level to directly overhead, and the leaf arrangement transitions into the seeds on the flower head. The topic brief did not ask for this step — Opus 5.5 planned it on its own.

Phyllotaxis opening comparison: GPT 6.1 design first, then Opus 5.5 design. Both coded by GPT 6.1.

4. Coding Ability Does Not Decide Whether a Video Looks Good

From our production experience, what determines the quality of a video is the blueprint. Right now, the blueprints GPT 6.1 produces are still a step behind Opus 5.5 in quality and visual appeal.

But GPT 6.1's coding ability is strong. It is fully capable of layer two. Once the blueprint is solid, swapping in a different model to write the code does not change the final video much.

5. A Finding That Goes Against Intuition

We cut roughly 20-second clips with matching content from both types of videos — one designed by Opus 5.5 and one designed by GPT 6.1 — and asked two video professionals to watch them blind: one editor with ten years of experience and one sound designer with twenty years of experience.

For the "counting spirals" pair, both picked the clip designed by GPT 6.1. For the other pair, the editor picked the Opus 5.5 design and the sound designer felt the two were about the same.

In every pair, both clips were coded by GPT 6.1, so the comparison is about design only.

This might suggest that GPT 6.1 is actually not bad at designing individual shots. The gap shows up at the full-video level — in the narrative: what to show first, what to show next, whether the viewer understands each step, and how to adjust if they don't.

A 20-second clip cuts all of this out, so these weak points do not show.

That said, we only asked two professionals and did a simple analysis. This should not be taken as a definitive conclusion.

6. Closing

First practical takeaway: the ceiling for Opus 5.5 is still high. The code-animated videos popular on X and Reddit right now have not yet pushed it to its limit, in our view.

Second practical takeaway: as of early October 2026, using Opus 5.5 for top-level planning and design is the best choice for making these videos. Coding and rendering can be handed off to other models.

We also tried a different coding combination along the way, and it did not work out. Here is what happened: in Google's Antigravity, we had Opus 4.6 write code following the blueprint from Opus 5.5. Halfway through, Opus 4.6 ran out of quota, so we let Gemini 3.8 continue in the same window to finish the code. The code was completed, but rendering kept failing: it ran for over an hour with no result. We asked it to fix things, and it quickly admitted the mistake and made changes, but the render still failed after a long wait. We eventually stopped. Since GPT 6.1 was already working well, we did not dig into why Gemini 3.8 failed inside Antigravity. The failure happened at the rendering step, so it does not say anything about the coding ability of Opus 4.6 or Gemini 3.8.

One last note: we are just a small content team. Our judgments come from hands-on creative work, not from AI engineering.