Choosing a Claude Model for Video
Model selection is useful only when the input, renderer and review standard stay the same. The public site supports credit-funded cloud generation, but does not expose live model comparison. In the local Opus test path, the configured model returns a title, duration and restricted Canvas animation code; older structured-storyboard ideas remain an evaluation reference, not a public workflow.
Compare a consistent input
Use the same product, three images, verified selling points and a 15-second target for every comparison. Record the model ID, request time and actual usage while keeping private material out of public logs. A model name or advertised capability is not a measured result.
Compare the output contract first
The current Opus bridge is evaluated as a bounded code response: title, duration and one Canvas frame function using only authorised asset IDs. A separate structured-storyboard route can be evaluated as scene data when it is explicitly enabled. Do not score a storyboard as though it were a finished video, or treat the Opus code response as a promise of live-action footage.
Score what people can actually review
Check that image references exist, the duration is within the brief, text fits and every claim matches the source material. Review the browser animation and the complete exported file with the same renderer and output settings. Ask reviewers to score readability, pacing and product accuracy on a five-point scale.
Record cost and latency
Keep the actual input, output and cache token counts, request latency and export time with the model ID. Calculate cost from the configured provider rates and observed usage, and label the rate date. Do not publish latency, conversion or quality scores before the corresponding calls and reviews have happened.
Current evaluation status
No public benchmark in this local build establishes that Fable, Opus, Sonnet or Haiku is the best model for a particular product. Administrator-only model comparisons use the configured provider account and its own limits. Public users can purchase credits and start cloud generation, but cannot select models or run this internal comparison. A future comparison should publish its input, sample size and review method alongside any score.
Choose the visual direction separately
Minimal, Bold Sale, Luxury, Tech Dark and Playful are rendering presets. Their fonts, colors and transitions are independent of a model comparison. Choose a preset that fits the product, then inspect the generated code or structured plan and the rendered result before deciding whether the workflow is suitable.