Opus 5.5: making short films and a full music video

On September 23, Deedy shared a roughly 26-second launch video for an inference service, built from paced typography, colored blocks and graphics. He reported about one minute and $2 for this attempt; we have not replicated those figures. Performance numbers inside the film belong to its copy and are not verified product measurements here.

His follow-up names JavaScript, Playwright and ffmpeg, without a video-generation model: create a browser animation, then capture and encode it. Opus organizes the explanation, arranges visual elements, writes code and uses tools. A video deliverable does not imply native generation of moving imagery.

That opens a useful route for product introductions, data stories and explainers. Code can precisely place text, numbers and shapes, while keeping names, colors and timing editable in the same animation.

• Watch Deedy’s original video ↗
• Read the author’s method ↗

• See Claude's official demonstration ↗The original release includes a Dashboards-to-Motion demo.

September's Opus creators wrote browser animations and captured them as video. On October 8, Anthropic introduced Claude Motion beta: turn reports, charts or product explanations into code animation, edit words, numbers and timing, then export MP4. An existing method is gaining a product entry.

Anthropic describes animation of text, shapes and existing images rather than video-model footage or generated people. A pricing explanation or staged chart differs from generating a person walking down a street. The official entry image shows Sonnet 5.5; this product release is not an additional Opus benchmark.

The accompanying Dashboards release connects data with inspectable queries and refresh times; Motion turns material into an explainer. Figures and narrative order still need checking before animation. Motion alone does not validate a report's conclusions.

Motion beta is available on Team and Enterprise: on by default for Team and requiring an admin switch for Enterprise. Docs, Slides and Design availability on Free, Pro or Max does not extend Motion to those plans. Start through /motion or the Artifacts template. We checked official instructions and imagery, without producing a Motion project in our own account.

Choose around the material you have: HyperFrames for source-level visual revisions; Manus for timeline and chat edits to code or supplied clips; Motion for eligible users explaining text and data through a template. This compares editing routes without asserting a shared model or production system.

• Revisit visual revisions in HyperFrames ↓
• See how Manus continues after a first version →

iArt’s public I only asked you to fix one line is a roughly 30-second square code animation. Asked to fix a typo before its user goes to bed, a robot repairs a brick, then rebuilds a city, reporting “+48,211 lines · 312 files” in the morning. Those numbers are part of the joke; the film and its source are both public.

An MIT-licensed HTML source accompanies the film. Canvas draws the picture, while Web Audio in the same file synthesizes the score and effects; timing, wording and staging remain editable. We also opened the original file in a browser and inspected frames at 6 and 28.9 seconds. iArt says the skills were built and tested with Opus 5.5; other models remain untested. Installation in another coding assistant does not establish equivalent results.

The example’s retrospective records why it changed: the first cut looked childish and placed text on objects. The next pass adopted a flat blueprint look, made each hammer blow dislodge the next brick, and restaged the desk, over-the-shoulder view and ending. This changes the causal story as well as the style. These are the maker’s recorded review passes. The repository also describes agent-made examples without human edits, so this is not evidence of manual human coding; no first-cut film is supplied for our visual A/B test.

For a small trial, use material you have permission to use in a 20–30-second explanation. Change one name, color or hold time, then check whether the same source preserves the rest of the picture and readable mobile text. This is our bounded trial suggestion, not a demonstrated client workflow. The open skill is free; model use and local tools still have costs. iArt’s web product charges by usage, offers a $2 signup credit, and requires a paid plan for MP4 export. This film’s actual costs, labor, customer payment and returns are undisclosed; product prices do not establish its cost.

The two routes also deliver differently. iArt’s method page, updated October 3, describes cloud-rendered exports for the web product: 1080p on Pro and 4K on Ultra. The open skill still produces HTML and MP4 on your own machine. For a 20–30-second film, the product page gives roughly 4–5 minutes and $0.60–$1.00 for generation; these are the maker’s estimates, not our measured time or cost.

The pricing page read on October 4 separates generation from export: roughly $0.25–$0.50 for 10 seconds, $0.60–$1.00 for 30 seconds, plus $0.09 per export render. One generation plus one export therefore totals about $0.34–$0.59 or $0.69–$1.09, excluding a paid plan, retries and human revisions. The existing animation’s actual bill remains undisclosed; these product estimates do not establish its production cost.

• Watch the maker’s 30-second animation ↗
• Open the animation’s HTML source ↗
• Read the first-cut issues and revision choices ↗

An editable animation also needs a way to inspect revisions. The October 1 Claude Code Mods tutorial demonstrates Replay Theater: a function rename produces five edits across three files, reviewed one diff at a time with /replay. The repository identifies Sonnet 5 for its demo and screenshots; the mod itself calls no model. This is an official open tool example, not a new Opus 5.5 capability or our animation retest.

Its code records edits when the tool is called, so a denied or failed edit can still appear. A displayed diff does not establish a successful file change. Each step shows at most 12 lines, and the in-session replay disappears on restart or plugin reload. It helps inspect intended changes; compare them with the actual files, execution results and full diff.

For the animation above, try one wording or timing change: inspect the code edits, then reopen the same work and check its text, picture and sound. This is our suggested review method; we have not installed the mod or run that experiment. The example requires Claude Code 2.1.287 or later. Mods execute with your permissions, so read the repository before installing.

• See the creator’s edit-replay example ↗

The HyperFrames desktop app brings chat, the picture and the timeline into one project. Its official example shows Claude Code and Opus 5.5 selected. Users can select an element, draw over the frame or attach a comment to a place and moment before sending a revision. This is official interface material, not a revision produced in our own test.

Marks first enter a pending-edit list and can be sent together after review. This connects a visible problem to a specific location or time range, rather than a description detached from the film. The timeline also supports manual splitting, moving and trimming, keeping human decisions in the project.

Official access currently covers Mac and Linux. The app requires HeyGen sign-in and a configured coding agent, whose account plan handles run usage. It is a tool for continuing code-based video work with different agents. Reviewing the resulting text, pacing and sound remains the way to judge whether a revision expresses what you intended.

• Compare timeline revisions in another product →The Manus demo offers another post-generation editing route; it does not establish use of Opus.
• See visual annotations and batched edits ↗
• Check desktop access and requirements ↗

Shinya Hara published a roughly 2-minute-10-second Jev explainer and a making-of account on September 24. He prepared an English shot plan with ChatGPT, then used Opus 5.5 in claude.ai, answering questions about duration and audio and making several small revisions. The 15-second request shown initially is a template, not the final film’s duration.

The useful lesson is specificity: when a concept appears, how elements move, how long text stays readable and how the next idea enters. Define one thing the audience should understand, then describe the visuals and pace so the model has a coherent explanation to build.

• Watch the complete Jev animation ↗
• Read the shot plan and revisions ↗
• Explore what Jev does →

• Watch the complete music video · 3 min 47 sec ↗Watch on the creator’s X page; login or regional restrictions may apply. Duration follows the author’s project notes; the nine frames below are a still preview.

CLAWDNAM STYLE reimagines Gangnam Style as a roughly 3-minute-47-second Clawd cartoon, moving from a sandbox to a subway, carousel and stage, with a reluctant horse recurring across scenes. Creator am.will (LLMJunky) published the film, role breakdown and project files. It shows code animation organizing a whole-song narrative and recurring characters, beyond a short typography reel.

In his September 28 account, Opus 5.5 handled research, writing, story planning and initial concept frames; Sonnet 5.5 continued from that material. The project assigns 14 song chapters to 14 parallel agents, then reviews and assembles them. It builds on John Heibel’s ClaudeAnimationBase painting engine, so model work, creative direction and existing tools all contribute.

The creator estimates about five hours across both models, including roughly two for Sonnet. He explicitly notes that Sonnet inherited Opus’s reference frames and code. This documents a design handoff, not equal results from independent starts; the two hours do not establish a time saving against Opus.

The published retrospective records a useful failure: a subway transition matched the leads and camera, but separate agents invented different passengers and train speeds, producing a visible jump. Its fix passes every visible character’s position, scale and pose, plus the scene state, into the next chapter. Smooth continuity requires a shared world as well as parallel workers.

• Read the actual Opus–Sonnet handoff ↗
• Explore the project and production lessons ↗

Aaron Orelup’s running account, dated September 28 and read September 30, now lists ten music videos: one fresh run per model at five effort levels. Each brief requests a 30-second music video, leaving the subject and method open. These are different creative executions, not matched-storyboard quality tests.

The ten video runs used the same machine’s GPU one at a time, with ComfyUI image, video and song models, Blender and the author’s tool notes available. Claude chose a route, operated tools and assembled the work; it did not natively output a video.

Production methods, music and timing below are the author’s account. Costs price recorded usage at API rates; the creator uses Max, so these are neither actual charges nor our replication.

Anthropic’s effort documentation provides context: effort affects thinking, responses and tool calls; lower settings generally use fewer calls. It is not a fixed token budget. Available tools and the amount of exploration also shape the result.

Our suggestion is to define the deliverable first: moving characters or camera moves over stills, a sung track or simple accompaniment. Compare versions that meet those requirements, recording waits, revisions and usage. The cost relationship changes between these settings; inspect the work delivered before choosing a model.

• Open Opus Medium · Last Tram to the Moon ↗
• Open Sonnet Medium · Moth Song ↗
• Open Opus Max · The Terrible Snail ↗
• Open Sonnet Max · All at Once ↗

Anthropic positions Sonnet 5.5 as a faster, lower-cost companion, with Opus 5.5 handling complex work that calls for careful judgment. The music video gives that distinction a practical form: settle the story, visual direction and difficult parts before delegating bounded chapters. For a product introduction or lesson, try a representative scene first, then decide whether another model should extend it. This is our suggested experiment, derived from the case.

On September 29, oxmannnn shared a 30-second Reddit motion-graphics reel: five builders worked on shots, five separate reviewers checked them, then fixers addressed problems. The creator reports rendering frames for the model to inspect, catching missing text, overflow and a crash at a particular frame. Visuals and synthesized audio share a timing schedule. These are the creator’s reported methods; we have not rerun the project.

The cost account is equally instructive. Using Max 20x, the creator received no API bill for the film. He priced the combined usage of the 15-second draft and 30-second version at $35.40 using Sonnet rates, then repriced those same tokens at $50.48 using Opus rates. The second figure is not an actual Opus run, and complete production costs were not disclosed. It cannot establish a saving at equal quality.

Independent evidence adds context: on September 28, Artificial Analysis reported an Intelligence Index of 56 for Sonnet 5.5 max versus 58 for Opus 5.5 max, with roughly 60% more output tokens per task for Sonnet. It used a pre-release deployment and disclosed a structured-output issue and plans to rerun relevant tests. This does not measure video aesthetics, but reinforces the distinction between lower token rates and lower task cost.

• See the 30-second reel and review method ↗
• Read the creator’s usage calculation ↗
• Continue with Sonnet 5.5 for everyday work →

The public diggerhq/shipvideo project connects the steps: accept a URL or prompt, extract page text and styling, ask Opus 5.5 for an HTML animation, render it in a browser and encode MP4 with ffmpeg. It checks script errors and text at selected times. The documented target is 20–40 seconds; this is a review of the implementation, not our own service run.

When another developer packages a successful experiment into a tool, the capability moves closer to everyday creation. Individuals and small teams can try clearer presentations of existing product pages, lessons or event material, while retaining responsibility for ideas, facts, selection and final review.

These examples extend from motion graphics to character-led music videos, with imagery still drawn primarily in code. Photorealistic acting or landscapes require further combinations of footage, image or video models, sound and editing. Start with the kind of imagery you need, then identify the model’s role.

• Explore shipvideo’s public workflow ↗
• Explore a fuller AI film production →

The works above both turn written requests into playable video, but take different routes. Deedy used model-written browser animation, then captured and encoded it; the Jev explainer involved the creator’s shot plan, answers to follow-up questions and revisions. The films depend on model-written code and tool use as well as human choices.

Released September 22, Opus 5.5 is Anthropic’s new flagship. The company emphasizes agentic coding, computer use and knowledge work, claiming performance near Fable 5.1 on most work. Those capabilities help explain its role in writing animation and operating tools, while vendor tests still need to be read separately from independent results.

Read September 25, the Artificial Analysis model card lists Opus 5.5 max with default fallback at 58, first among its 211 comparable models. That is a reason to examine it closely, within the stated evaluation, effort setting and runtime conditions.

GitHub reports comparable task resolution to Opus 5 with fewer steps and tokens in early testing, plus recovery from errors in multistep work. Copilot integration brings those model changes into tools developers already use.

Short, well-defined work deserves its own comparison. In a personal benchmark updated September 30, developer Ibragim Badertdinov selected ten engineering tasks and ran each configuration three times per task. Using Pi 0.85.1 at high effort, Opus 5.5 and Sonnet 5.5 both passed all 30 attempts. The table lists per-task cost and time of $0.3034 / 102 seconds and $0.0971 / 46 seconds respectively. These are API costs, not subscription allowances. Earlier Sonnet 5 passed 25 attempts in the same suite, providing a version reference.

Several models already pass every check in this small suite, and the author warns that equal scores do not imply equal code quality. The complete tasks remain private, and we have not obtained full traces for independent review; the method describes a simplified JSONL importer with validation and atomic replacement. Alongside the longer tasks and music videos above, it illustrates why scope matters when switching: try Sonnet on a familiar bounded task, recording results, waits and revisions. A whole film or ongoing project still needs comparison against its own deliverable.

• See the personal engineering benchmark and settings ↗

Microsoft and Hugging Face's October 3 report adds a different test of agent work: what is left in the database. In a synthetic delayed-delivery case, an agent made nine tool calls, then marked the support ticket solved while the shipment exception remained unresolved. The required state was hold, pending the carrier's response. The task and failure trace are public. The paper uses this case to explain grading; it is not identified as an Opus 5.5 trace.

ThinkingBox repeats each of 507 business tasks twenty times from clean starting states. Average attempt success, at least one success in twenty, and success in all twenty answer different questions. Opus 5.5's average rises from 66.50% to 67.16%; both versions still pass all twenty on 241 tasks.

These are synthetic workflows reported by the authors, not our retest or a production guarantee. Coding checks and video quality measure different outcomes. When an agent handles an order or application, inspect the changed records, unresolved steps and unintended changes alongside its reply.

• Read the authors’ tasks, results and method ↗
• Inspect the synthetic task’s required hold state ↗

The opening Opus 5.5 medium image and this Opus 5 medium image come from the author’s shared-task grid. Two distinct legs are visible in the newer output; one dominates the older image. Their styling also differs. This describes these particular images, not every drawing either model will produce.

The author also reports two max-effort attempts that exhausted the output allowance without producing an SVG. More reasoning does not guarantee completion of this task. A practical starting point is medium effort, raising it for a specific problem while retaining outputs for comparison.

• Open the complete comparison grid ↗Original SVGs, usage and failed attempts across models and effort levels.
• Read the author’s account ↗An independent author’s experience, not a BitShovel test.

Developer Senko Rašić gave different coding agents the same request for an explorable voxel game in one HTML file, then published the browser outputs and source. The September 22 Opus 5.5 page shows 3D terrain, world settings and a save control. His collection also lets readers open the July Opus 5 and September Fable 5.1 and Astra versions. We inspected the opening screens of all four, not their complete gameplay.

These works make “good at coding” concrete: movement, mining, placement and inventory must become a running interaction. For this one Claude Code run at high effort, the creator reports about 50 minutes and $19 for Opus 5, versus 48 minutes and $14 for Opus 5.5. The runs took place on different dates, so this is not an average saving. The same-task Astra and Fable entries do not list costs, making a four-model value ranking unwarranted.

A fixed request controls the task but not the coding agent, date, dependencies or randomized world. Open the results, check whether controls respond and inspect the source before judging reliability; a polished landing screen or one run’s duration cannot settle it. For readers making a webpage, tool or game, these inspectable outputs offer a better starting question than an abstract score alone.

• Continue: one game from prototype through repeated revisions →
• See the four versions and fixed prompt ↗
• Open the Opus 5.5 browser result ↗
• Compare the Opus 5 result ↗

• Open the original shared-task 3D works ↗Choose Landmarks and real places → Eiffel Tower, then select Opus 5.5 or Muse Spark 1.3.

The voxel games above show a request becoming a webpage. AA-3D adds a more consistent reference: 55 scene briefs use the same Stirrup agent framework and offline tools, then recruited panelists interact with two anonymous outputs for the same task and choose a preference. It measures relative preference for these works, not reliability across all games or the visual quality of video-generation models.

On the original site, we opened Opus 5.5 and Muse Spark 1.3 outputs for the same Eiffel Tower brief. Both displayed the tower, its surroundings and floor controls. In this visit, Spark's camera altitude changed from 62.1 to 118.5 metres when selecting Level 1 then Level 2, and its nearest-level label changed too. Opus rendered its overview, but we did not confirm a viewpoint change through its first- and second-floor controls. That is an unconfirmed step in this browsing environment, not a finding that every user cannot operate it. The Opus image uses the site's HUD switch to hide the interface, while Spark retains its UI; these are not camera-calibrated views for visual scoring.

Keep the overall score separate from this one-task observation. In the source table checked October 15, Opus 5.5 scores 1190 and Sonnet 5.5 scores 1194; their 95% intervals overlap, so four points do not establish a decisive winner. Meta's entry here is Spark 1.3 at 1035. Wonder and Wonder Full have no entry; neither Spark's 3D result nor a separate general text index should be attributed to Wonder.

For a knowledge exhibit, webpage or small game, a useful sequence is to run the artifact, try one explicit action, check that the scene and feedback change together, then examine detail and cost. AA provides three.js, with optional Rapier, Blender and an asset library: tools that turn model-written code into a work. In your own scene, an assistant can revise the controls, camera and rules in code before you reopen it to check. This visit inspected original artifacts, without regenerating them or validating a complete game.

• Read AA-3D's method and source table ↗
• Continue with Wonder's four public build comparisons →A different creator and agent setup; the results are not combined.

Anthropic emphasizes clearer communication: leading with the important point, using less jargon and following writing instructions. Its release page includes paired answers readers can inspect. These are vendor-selected examples.

• Compare the official answers ↗See the Communication section of the release.

The official cost guide lists standard API prices per million tokens: $4 for fresh input, $20 for output and $0.20 for cache reads. Input and output rates fall 20% from Opus 5, cache reads 60%. Subscription allowances and API bills are different measures.

A task’s bill also depends on turns, cache hits, reasoning output and retries. Compare the same material and completion requirements, then the result and actual bill; a rate reduction alone does not establish the saving on every task.

Existing Copilot users can check the model picker on supported plans and clients. GitHub describes a gradual rollout, with organization access affected by administrator settings. Claude users should check their account’s model and usage controls first.

• Sonnet 5.5: how should everyday work be split with Opus? →
• Read the cost guide and calculator ↗Distinguish illustrative calculations from your actual usage.
• Check Copilot availability ↗Plans, clients and organization settings.

If you use Opus 5, compare a familiar task: does the same material require fewer clarifications or revisions, and is the final explanation easier to assess? If you prefer GPT, Grok or GLM, try one task your current tool struggles with and compare complete outputs. You can add a useful option while keeping your established workflow.

Astra’s 3D scenes, the code animations here and generated footage in AI films represent different creative routes. Pick the result you want, then compare the process, editing flexibility and costs. This topic will continue into Chinese explainers, public source files and cross-model productions.

• Explore real Astra creations →From a description to an editable 3D scene.
• Compare Sol, Luna and Opus costs →Separate rates, task bills and subscriptions.
• Beyond model performance: Duet’s ongoing work and approvals →