San Francisco-based Palette Technology wants creative teams to stop choosing which AI video model to use. The Y Combinator-backed startup unveiled a platform in late July 2026 that makes that decision for them, routing projects across eight different models—Google Veo 3, Kling 3 Pro, PixVerse, and others—depending on what each job requires.
The pitch is straightforward, if ambitious: consolidate video generation, editing, and workflow management into one interface that figures out which tool works best, then gets smarter with each project. Whether that solves a real bottleneck or just adds another layer to an already crowded market is the question Palette now has to answer.
Josephine Lee and Serena Pei, both MIT computer science graduates who spent time at Scale AI, Datadog, and Amazon, started the company earlier this year. They're entering a field where Runway, Luma Labs, Pika, and others have already staked claims, most tethered to proprietary models. Adobe plays in the space too, though its Firefly ecosystem leans toward enterprise creative suites rather than a unified generation layer.
Palette's approach differs in the routing. The system analyzes incoming requests and assigns them to whichever model fits best from its roster: Veo 3, Kling 3 Pro, PixVerse V5.5, Seedance 2.5, MiniMax H3, Flux 3, Wan 3, or Google Omni Flash. When a job fails or a less expensive model could have handled it, the platform refunds credits or returns the difference.
The company claims the method cuts generation costs in half and speeds output by 35% compared to single-model setups. Those figures come from internal benchmarks run against 130 automated tests, so the usual caveats about self-reported performance apply.
One feature Lee and Pei emphasize is character consistency. The platform keeps visual identity intact across multiple shots and formats, something that matters when a brand needs the same animated spokesperson in twenty different social clips. The website showcases eight character sheets—erika, juliette, kagemi, zlatia, shigure, suseka, jane, sasimi—demonstrating how appearance holds steady through scene changes.
Users control targeting, format, pacing, and visual direction through a storyboard interface tied to the routing layer. Natural-language commands handle scene restyles, object swaps, and sequence edits without tool-hopping.
Palette charges a cent per credit for individuals. A thousand credits buys a full character reel: sheet, storyboard, and fifteen seconds of character-consistent video. Enterprise customers get access to what the company calls an "Autonomous CMO" service, along with ad generation, brand-safe user-generated content tools, and learning modules. Palette declined to share enterprise pricing publicly, pointing prospects toward custom quotes instead.
Developers can tap the same features through a REST API at studio.palettelabs.com/api/v1, drawing from the same credit pool as the web interface.
The startup participated in Y Combinator's standard half-million-dollar investment structure. Palette's homepage displays backing logos for both Y Combinator and OpenAI, though the nature of OpenAI's involvement remains unclear.
"Unify the creative workflow into a single, high quality generation engine," Lee wrote in the company's Y Combinator launch post.
Palette's privacy policy, updated August 22, 2026, states customers own their outputs and that the company won't use customer data or generated content for model training. Agreements with model providers supposedly include the same restriction. SOC 2 Type II and ISO 27001 certifications are on the roadmap but not yet secured, according to a security page last updated August 7, 2026.
The platform's timing is interesting. As video generation models proliferate and costs drop, the friction shifts from "can we make this?" to "which tool should we use?" Palette is betting that answering the second question automatically has value. Whether creative teams want that decision made for them, or prefer keeping their hands on the wheel, will determine how much traction the startup finds.
