The tooling was there. That wasn't the issue. By early 2026, if you wanted to generate video with AI, you had options—Runway, PixVerse, Google's interfaces, a rotating cast of models with names like Seedance and Hailuo that sounded vaguely like wellness apps. Josephine Lee and Serena Pei, both MIT computer science grads, watched marketing teams and developers ping-pong between platforms, toggling subscriptions, writing custom API wrappers for each new release.
What didn't exist, they figured, was the layer in between: something to route the work, manage the mess, and spare creative teams the cognitive overhead of choosing which model to fire up for which job.
That observation became Palette, a YC-backed startup that went live last summer with a pitch that sounds almost mundane until you consider the fragmentation it's addressing. The platform is, in essence, an orchestration canvas—users brief a project, and Palette routes requests to whichever model in its roster makes sense. Google Veo 3.1 for one task. ByteDance Seedance 2.0 for another—though availability of the latter may be limited due to a paused global rollout. Kling 3.0 Pro, MiniMax Hailuo 02, PixVerse V5.5, LTX Video. You get the idea.
Pick manually if you want. Or let Palette decide based on cost, speed, and what the company describes as "100% task-capable routing" according to its internal benchmarks—a claim that, like much in this space, hasn't been stress-tested by outside auditors.
One Canvas, Six Models, No Spreadsheet
Lee, now CEO, and Pei, the CTO, describe Palette as "the creative engine for multimodal content," which is the kind of phrasing that tells you exactly nothing until you see the interface. The idea is to consolidate what currently requires multiple subscriptions into a single workspace. Text, images, video, music. Brand kits. Character consistency across shots—something that's harder than it sounds when you're stitching scenes together. Real-time collaboration baked in.
The workflow problem is real enough. A 30-second character-consistent video might mean toggling between three platforms, exporting files, importing them elsewhere, hoping the style transfer holds. Palette's answer is to handle generation, editing, storyboarding, and asset management in one place, accessible via web or API. There are modules for performance direction and auto-repurposing, though the company's website is light on granular details about how those features actually work in production.
Palette claims its routing delivers 35% faster generation and 50% lower costs compared to using individual APIs directly. It runs more than 130 automated regression tests, the company says, and refunds credits automatically when jobs fail. If a task gets routed to a cheaper model than expected, users pocket the difference. These are the kinds of operational details that matter more than the pitch deck promises—if they hold up at scale.
The Economics (With a Few Asterisks)
Pricing starts at $0.01 per credit on the individual tier. Palette suggests 1,000 credits can produce a full character reel: character sheet, storyboard, 15 seconds of consistent video. That math hasn't been validated by third parties, and credit-to-output ratios in generative AI tend to be squishy depending on resolution, complexity, and how many times you hit "regenerate."
Enterprise customers get custom quotes for what Palette calls "end-to-end services" tied to product launches and campaigns. Developers, meanwhile, can tap API endpoints for images, video (async jobs), music, upscaling, vectorization. Authentication is bearer token-based. Jobs can be polled, models listed, assets managed or deleted—standard infrastructure, presented without much flair.
The company went through Y Combinator's Summer 2026 batch and is based in San Francisco. Serena Pei's personal site states Palette raised a $500,000 pre-seed round led by YC in February of last year, though this has not been confirmed independently. Neither founder has publicly disclosed investor details beyond the accelerator. The team, per YC's directory, is just the two of them.
A Market Expanding Faster Than the Tools Can Keep Up
The AI video generation market is in one of those early growth spurts where the numbers vary wildly depending on who's counting. The Business Research Company pegs the market at $850 million in 2025, projecting it'll hit $2.07 billion by 2030. Grand View Research is slightly more conservative: $788.5 million last year, climbing to $946.4 million in 2026 and eventually $3.44 billion by 2033.
Either way, the trajectory is steep. And the space is crowded in a way that makes naming collisions inevitable. Palette shares its name with at least three unrelated products: Spectro Cloud's PaletteAI (launched last July for AI infrastructure management), SiMa.ai's Palette Neat (a Physical AI environment from last June), and a Korean video tool at pltt.ai. None are direct competitors, but it's the kind of namespace collision that makes branding in AI feel like a land rush.
Palette's website lists logos for Y Combinator and OpenAI under a "Backed by" section, though the OpenAI association is unverified and lacks independent confirmation. It's not uncommon for startups to showcase association in ambiguous ways—though transparency around such relationships would help.
The Shifting Ground of Model Availability
Then there's the model ecosystem itself, which remains something of a moving target. ByteDance's Seedance 2.0 launched officially in February 2026, but TechCrunch reported a month later that the company had paused its global rollout—potentially limiting access depending on region. Google's Veo 3.1 became available via the Gemini API and AI Studio in January. LTX released open weights for LTX-2.3 in March. Runway deprecated its Gen-3 Alpha and Turbo models last July as newer versions shipped.
Palette's value proposition hinges, in part, on maintaining access to this rotating roster of models. If one goes offline or restricts API access, the routing logic has to adapt—or users are back to managing fallback plans themselves.
Competitors are circling similar ideas. Shutterstock launched an AI Video Generator in April 2026, aggregating models under a commercial-ready wrapper. AKOOL announced its Agentic Canvas in June, also targeting enterprise workflows with model abstraction. Both are chasing the same thesis: that teams generating content at volume will pay for a unified control plane rather than cobble together API calls and pray nothing breaks.
Betting on the Orchestration Layer
Palette's product went live with updated security and trust documentation as of late June last year. Some of the legal pages include placeholder language—inconsistencies the company attributes to starter templates provided for convenience. It's a minor detail, perhaps, but the kind that makes you wonder how much of the infrastructure is still being figured out in real time.
For now, Lee and Pei are betting that orchestration is harder than it looks. That the teams churning out hundreds of videos a month—marketing agencies, content studios, in-house creative ops—will eventually tire of managing six subscriptions, debugging API changes, and second-guessing which model to use for which brief.
Whether Palette becomes the default layer for that workflow, or gets subsumed by a larger platform with deeper pockets, is the kind of question that usually resolves itself quietly over 18 months. The founders have a head start. But in a market this fluid, being early and being right are not always the same thing.
