Every month, something new lands in the developer ecosystem. Most of it is noise. Some of it genuinely changes what you can build in a hackathon weekend. This is the filter.
The Standard Still Holds
Before the new stuff: the core stack that wins hasn't changed. Next.js + Supabase + Vercel AI SDK remains the dominant pattern across winning projects. It deploys fast, covers auth + database + AI in one coherent system, and judges have seen it work. Don't let new tools distract you from this baseline if you don't already have it locked in.
What's Actually Worth Trying This Month
Vercel AI SDK 4.x — Streaming is now trivial
The SDK's generative UI capabilities have matured to the point where streaming AI responses into React components is copy-paste simple. The useChat hook handles loading states, streaming, and error boundaries automatically. If you're adding any conversational feature to your project, this is your starting point — not building your own fetch loop.
Hackathon fit: High. Setup is under 15 minutes. Judges see streaming text as more impressive than a loading spinner even when the underlying model is identical.
Pitfall: Don't use it for batch processing or non-conversational tasks. The streaming pattern is optimized for chat-style interactions.
Claude claude-haiku-4-5 via Anthropic API — The speed-cost sweet spot
For classification, tagging, summarization, and structured extraction tasks, Haiku is now fast enough and cheap enough to use as a real-time processing layer — not just as a demo feature. In the health, civic, and legal domains, where you need to handle sensitive content carefully, Claude's constitutional training is a meaningful advantage over alternatives.
Hackathon fit: High for content-heavy domains. Use for document parsing, intake forms, translation, and routing logic. Reserve Claude Sonnet / Opus for the user-facing moments that matter.
Pitfall: Don't use Haiku for complex reasoning chains. It's excellent at single-turn structured tasks, not multi-step agentic work.
RAG without a vector database — Simpler than you think
The dominant RAG pattern in 2025 involved setting up a vector database, embedding pipelines, and retrieval infrastructure. For a 24-hour hackathon, that's too much to manage. A lighter pattern has emerged: embedding documents at startup, storing them in memory, and doing cosine similarity in JavaScript. For datasets under ~1000 chunks, this works fine and eliminates an entire infrastructure dependency.
When to use it: You have a fixed corpus — legislation, medical guidelines, a company's documentation, a city's open data — and you want an AI to answer questions about it. This pattern gets you there without Pinecone, Weaviate, or any managed vector service.
Concrete approach: Vercel AI SDK's RAG guide + a simple in-memory embedding store using their embed() function. The whole pipeline is about 60 lines of TypeScript.
TanStack Query v5 — Data fetching that doesn't fight you
If your project has any server-side state — user data, API results, real-time feeds — TanStack Query v5 is worth adding. The devtools alone have saved projects during live demos: you can inspect cache state, force refetches, and debug data issues without console.log archaeology.
Hackathon fit: Medium-high. Add it if you have more than two or three API calls. Skip it if your data flow is simple.
What to Skip This Month
Bun for production — Fast locally, still rough edges on Vercel deployments. Use Node. Not worth debugging build issues at 11pm.
LLM agents for core features — Agentic loops are impressive in demos and unreliable in live judging. Build the happy path first; agents are a bonus if you have time left over.
New database options — Supabase is still the best free-tier option for hackathons. Don't let "Neon vs Supabase vs PlanetScale" become a decision you're making at midnight.
The Principle
The best hackathon stack isn't the newest or most technically impressive — it's the one you can get working in the first two hours so you can spend the rest of the time building the thing that actually matters. Every new tool you add is a bet that the integration time is worth it. Most months, it isn't.
This month, the Vercel AI SDK update is worth it. Everything else on this list is conditional on your domain.
What's in your stack this weekend? We're tracking winning project stacks across hackathons — if you built something this month, add it to the showcase.