How to make videos with Claude Opus 5.5: workflows, prompts and costs

Claude makes videos by writing code that draws every frame, meaning the visual pixels come from a browser, Blender, or a local renderer rather than a video generation model.

Anthropic officially launched Claude Motion on October 8, 2026, describing it as an in-product feature that writes code to animate typography, statistical charts, vector shapes, and user-supplied images. According to Anthropic, the feature allows creators to alter any word, numerical metric, or timing value directly in an editor, and says Claude Motion does not use a video generation model: no generated footage and no AI-generated people. Finished animations download as an MP4. The beta remains restricted to Team and Enterprise workspace tiers.

Local rendering engines follow this identical programmatic principle. The official HyperFrames repository explains that its renderer seeks each frame in headless Chrome and encodes the frames with FFmpeg, so the same input always produces the same video. The Remotion framework operates through the exact same philosophy by rendering React component states to disk frame by frame.

This programmatic reality is reflected across developer communities on X, where 344 posts in our collection explicitly stated that their published video consisted entirely of code without relying on generative video algorithms. At the same time, only 1,772 posts, or 24 percent of the 7,406 total posts analyzed, identified any specific software library or framework by name, as the vast majority of posters simply attributed the output broadly to Opus 5.5.

Anthropic officially released Claude Opus 5.5 on September 22, 2026, with input and output prices 20 percent below Opus 5. Our dataset spans from September 20 to October 17, 2026, capturing 7,406 posts from 3,729 unique accounts, with 96 percent of those posts including embedded video files demonstrating these workflows.

In the editor's reading of this architecture, producing videos via code provides complete programmatic determinism and granular version control, allowing developers to change a single layout dimension or timing offset and re-render the composition without random generative drift. The corresponding trade-off is that any project demanding photorealistic live-action footage or realistic human characters still requires an external video generation model.

Which route fits depends on your plan and on whether you need motion graphics or realistic footage.

Table: Four ways to make a video with Claude

Sample/method: Route details from Anthropic, HyperFrames and Remotion official pages (checked 2026-10-18); counts are rule-based matches across the 7,406 posts in this collection, including negative mentions, and measure discussion only.

The first route is Claude Motion inside Claude: you prompt, adjust in the editor or ask Claude to change it, and download an MP4. Anthropic says Motion projects can be opened in Adobe, Descript, HeyGen, Higgsfield, invideo, Luma AI, and Runway. Enterprise admins must turn Motion on in Organization settings > Artifacts.

Charlie Hills (@charliejhills) shared a grid of 16 Apple-style Claude Motion graphic sequences on October 13, 2026, explaining that giving it his brand kit first got him more of the same, so he rewrote the brief around one bar, Apple quality, and now changes only the one line that describes the scene. Claude Motion is available only on Team and Enterprise plans for now. (Item: How Claude Motion produced 16 Apple-style animations)

The second route pairs Claude Code with Remotion or HyperFrames. According to Remotion documentation, you create a blank project with npx create-video --yes --blank my-video, add the skills with npx remotion skills add, start npm run dev, then run claude in the folder and describe the video. HyperFrames (Apache 2.0, from HeyGen, no per-render fees) installs as a Claude Code plugin, needs Node.js 22+ and FFmpeg, and renders plain HTML with npx hyperframes preview or render. Its README notes that Remotion uses a source-available Remotion License.

Chubby♨️ (@kimmonismus) reported producing a three-minute historical video about artificial intelligence on September 23, 2026, using roughly 7,400 lines of React and TypeScript written entirely by Claude Code inside Remotion with vector assets generated via SVG and Canvas. The project required approximately one hour of execution time and consumed 7 percent of his weekly account quota without stock footage or image generators, according to his post. (Item: Opus 5.5 renders a 3-minute AI history film entirely in code)

Jake Moran (@jake11moran) posted a clip on September 22, 2026, showing a 3D camera move and the 3D text "it writes the script," made with Opus 5.5 and HyperFrames. It shows code-driven text graphics layered over live footage. (Item: 3D camera moves and text effects with Opus 5.5 and HyperFrames)

The third route is plain web code or Blender scripts. Three.js/WebGL is the most-named tool in the collection (539 posts), Blender has 325. Many of these "videos" are screen recordings of interactive pages rather than rendered films.

Kevin Ngo (@kevin_t_ngo) demonstrated an animation on September 30, 2026, showing a boxy little robot playing a piano next to a black cat. He says Opus 5.5 wrote the piano music in Python, created the 3D animation in Python, and rendered it in Blender. (Item: Opus 5.5 writes piano music in Python and renders a 3D animation)

The fourth route is a hybrid. Across the collection, 378 posts name a video or image generation model such as Seedance or Veo. In these setups Claude writes the motion graphics, the edit or a 3D blockout, and the generation model supplies the footage.

A.I.Warper (@AIWarper) presented a retro anime sequence on October 3, 2026, with the finished shots above and the grey Blender blockout below. His listed stack: Opus 5.5 with Blender MCP, Seedance 2.5 (a video generation model), and DaVinci Resolve Studio driven by Opus 5.5. (Item: Opus 5.5 drives Blender and DaVinci Resolve for a retro anime short)

Sample/method: all 7,406 posts in this collection (2026-09-20 to 10-17, UTC+8), rule-based match on tool names in the post text, including negative mentions such as 'no Blender'; only 1,772 posts name any tool.

After 10+ hours of tuning, X user Deedy (@deedydas) shared tips we did not test: use Claude Code, not the app; Manim, HyperFrames or Motion Canvas for motion graphics; an image model for keyframes and Veo 3.1 or Seedance 2.5 for footage; reference images for consistency; an animatic before the full video; OpenTimelineIO to assemble.

Among the 1,985 posts in our classified sample showing an author's original work, narrative stories, music videos, and cartoons represent the largest category with 391 examples, followed by 339 3D scenes and 289 product promos or ads.

Sample/method: random sample of 2,500 of the 7,406 posts in this collection (2026-09-20 to 10-17, UTC+8), 1,985 of them the author's own work; AI classification with 30 hand checks. Counts measure discussion volume.

Motion graphics, kinetic typography, and user interface showcases accounted for 236 posts, games comprised 229 posts, and technical explainer videos or dynamic data visualizations totaled 185 posts. The remaining projects fell into generalized creative experiments or unclear formatting.

As an example of long-form narrative production, creator Ziwen (@ziwenxu_) reported producing a complete 25-minute animated episode titled Gus & Spark without traditional video diffusion models. According to his post, Opus 5.5 max worked on it for over 15 hours and wrote 34K lines of code, consuming 986 million tokens at an estimated API price of roughly $370.

Because our sample captures public posts on X, showy 3D visual experiments, game demos, and stylized music videos may be overrepresented relative to private corporate presentations, dashboard animations, and internal product explainers.

Creators who regularly publish functional video code converge on three distinct prompting behaviors: specifying technical constraints up front, providing an explicit aesthetic reference to emulate, and locking down a static brief while varying only a single scene parameter.

To ensure consistent browser rendering, developers frequently define the precise technical execution environment. On September 29, creator dawood46 (@notdwd) shared a prompt instructing Claude to generate a 12-second motion sequence at 1920 by 1080 resolution and 30 frames per second, packaged within a single HTML file where every animation tick is governed by a seek function rendered via Playwright, FFmpeg, and programmatic audio synthesis. The author offers a commercial development playbook based around this structure.

Providing concrete visual references prevents the model from defaulting to uninspired visual tropes. On September 28, dawood46 (@notdwd) warned that asking Opus 5.5 to invent animations from scratch causes it to fall back on generic compositions. Similarly, Archive (@ArchiveExplorer) noted on September 29 that without structural references, Claude's default design pattern is centered text, a gradient background and everything fading in. Archive recommends giving it a brand brain first: logo, colors, fonts and screenshots of your real product.

Maintaining a fixed structural template simplifies iteration across multi-scene projects. Charlie Hills (@charliejhills) found that giving it his brand kit got him more of the same, so he rewrote the brief around one bar, Apple quality; the brief never changes and he rewrites only the line that describes the scene.

For product walkthroughs, simple functional constraints can also succeed. Artem (@sl1ma4) quoted his baseline prompt directing Claude Code to create a dynamic presentation video about a specified project using Three.js, asking for it to be "not too fast" so people "should actually understand what it's about."

In addition, Deedy (@deedydas) suggested adding negative prompt instructions to discourage repetitive model habits, such as short punchy sentences or excessive numbers, advising users to instruct Claude to narrate like a university professor.

Among 245 posts in the collection that explicitly stated what Claude cost for a video or project, the median reported expense was $10.75.

Sample/method: 245 posts in this collection (2026-09-20 to 10-17, UTC+8) stating a dollar amount actually spent on Claude; 626 rule-matched candidates checked by a strict AI pass, amounts extracted by rule with 20 hand checks, relays of one project counted once. All self-reported, mostly estimated at API prices.

Table: Claude 5.5 API prices (USD per million tokens)

Sample/method: Anthropic's Opus 5.5 announcement (2026-09-22) and Claude Platform model docs, checked 2026-10-18; the Batch API is 50% off.

Reported figures split into two primary distributions. A group of 95 same-prompt model benchmark tests showed a median cost of $9.94, with 58 of those runs falling tightly between $5 and $20. In contrast, 150 diverse multi-step production projects exhibited a wider spread with a median of $14.02, including 26 complex initiatives that reached $100 or more in self-reported API costs.

Ryan Sael (@RyanSael) shared an interactive educational demonstration on September 23, 2026, simulating camera lens elements and focal planes, which Opus 5.5 authored in a single run lasting one hour and 26 minutes at a self-reported API cost of $25.66. The output is an interactive web page captured as video rather than a traditionally compiled animation file. (Item: Opus 5.5 builds an interactive lens focus lab in one shot)

Artem (@sl1ma4) shared a Three.js product animation on October 5, 2026, generated from a single Claude Code prompt describing an application monitoring service, which completed after one hour and seven minutes at an API cost of $12.45 across 185.5k output tokens and 1,767 added code lines using Opus 5.5. In an earlier test on October 1, the same author generated a 30-second motion graphics clip for Uptimepage using Sonnet 5.5 in 26 minutes for $6.74. (Item: One prompt has Opus 5.5 build a presentation video in about an hour)

Selecting smaller models can lower token expenses significantly. Benchmark testing from Bridgebench (@bridgebench) on September 28 revealed that the same lava lamp prompt cost $1.27 across 9 minutes and 35 seconds on Opus 5.5, compared to $0.47 across 5 minutes and 44 seconds on Sonnet 5.5, and the author said Sonnet nailed the glow, the lighting and the details.

Large-scale computational projects occupy the high end of reported expenditures. Stefan 3D AI (@Stefan_3D_AI) reported spending $2,175 in Opus 5.5 tokens over 36 hours and 7.85B tokens, with access to Unity, Blender and APIs, plus $230 for 2D, 3D and music generation, and getting a playable old-school MMO RPG. While such massive software undertakings are rare in the dataset, they illustrate the upper boundary of open-ended agent execution.

Conversely, 20 posts in our cost sample reported an effective marginal cost of zero dollars because the authors relied entirely on standard consumer subscription allowances. For instance, Chubby♨️ (@kimmonismus) noted that compiling his three-minute historical Remotion film required about an hour of time and consumed 7 percent of his weekly usage quota. In programmatic workflows, token generation represents the primary expense, as rendering tasks handled locally via HyperFrames or Remotion run on local compute hardware without per-render software fees.

As a methodological caveat, five specific accounts accounted for 38 of the 95 benchmark tests analyzed, primarily centered on a recurring prompt genre simulating soft 3D physics toys, meaning those test results capture a specific viral aesthetic rather than a comprehensive cross-industry baseline.

Direct critical complaints represent a minority of total posts, but they highlight key operational boundaries around plan tiers, token quotas, and repetitive design tropes.

Across our classified sample of 2,500 posts, 2,285 posts raised no structural criticisms, while 90 posts cited generic or buggy visual output, 74 posts noted token costs or strict usage rate limits, 46 posts emphasized that output files still required manual developer intervention, and 5 posts criticized plan restrictions.

Subscription limitations remain a friction point for individual creators. Anthropic officially restricts the Claude Motion beta to Team and Enterprise workspace accounts. In response, creators like James Malsawm (@JamesMalsawm) have publicly requested that Anthropic extend Motion functionality to Pro and Max subscription tiers. Concurrently, Mustafa Al Bassam (@musalbas) expressed frustration on October 15, saying Opus now uses the built-in Claude Motion by default instead of scripting from scratch (his experience; Anthropic has not confirmed it).

Commercial revenue claims, including mentions of client contracts or monthly business earnings, appeared in only 17 of the 2,500 sampled posts, representing under 1 percent of discussions. The editor did not verify any of these financial declarations, and they are omitted from our findings.