Descriptions:
Part 3 of Web Dev Cody’s AI video SaaS series delivers two major additions: multi-model video generation support and a testable service architecture. The episode walks through integrating three video generation APIs — WAN 2.2 (the existing baseline), Cling, and Google’s VO3 — complete with real pricing context: Cling charges approximately $0.35 per 5-second clip while VO3 runs around $0.75 per 5 seconds, making cost-aware model selection a practical concern when generating full videos.
A key engineering challenge covered is building a duration adapter pattern so the CLI’s high-level `–duration` flag maps correctly to each model’s internal parameter structure. WAN uses frames-per-second and frame count; Cling and VO3 use their own abstractions. The solution introduces a model-agnostic duration type resolved at runtime, keeping the CLI interface clean regardless of which backend is selected.
The second major refactor extracts the growing generate.ts file into separate service modules for OpenAI, ElevenLabs, FAL AI, and FFmpeg, then adds a Vitest test suite that mocks all external services to verify the happy-path generation flow. The video also demonstrates theme flag support — applying consistent anime, cinematic, or custom style prompts across all generated video clips. Developers building multi-provider AI media pipelines or studying how to structure agentic coding projects for long-term maintainability will find this episode particularly actionable.
📺 Source: Web Dev Cody · Published February 18, 2026
🏷️ Format: Hands On Build
