Arena.ai's Code Arena Now Builds and Ships Real Fullstack Apps

Code Arena, Arena.ai's platform for comparing AI coding models head-to-head, just made a significant leap. It started as a frontend prototyping sandbox where you could watch models build web UIs and vote on the results. Now it supports full-stack development: PostgreSQL databases, user authentication, third-party API integrations, and one-click deployments to Vercel. The pitch has changed from "compare models" to "build and ship real software."

What was missing before

Traditional benchmarks measure correctness: whether code compiles and passes a set of static test cases. Correctness matters, but it's only part of what defines real development. Building software is iterative and creative: you plan, test, refine, and repeat. Code Arena was already ahead of static benchmarks by letting models act as agents and build live apps. But the apps it could build had a hard ceiling: no login flows, no persistent data, no backend. That ruled out anything resembling a real product.

AI coding models are evolving faster than ever before. The question is no longer simply whether a model can write static code, but how well it can build real end-to-end applications. Fullstack Code Arena is Arena's answer to that question.

What's new

The update ships five concrete capabilities that weren't possible before:

• Database integration: A PostgreSQL layer that lets agents generate code for database schemas, user authentication, and Row Level Security (RLS) , a Postgres feature that enforces access control at the database row level, so users can only read or write their own data.
• Third-party API access: You can now securely pass in API keys to connect to external services , an OpenAI key for a chatbot, a Stripe key for payments, or any other REST API.
• Persistent dev server: A live server running inside the sandbox with hot reloading, so changes reflect instantly without a full rebuild.
• Bash and web search tools: Agents can now run arbitrary shell commands and search the web for documentation in real time, dramatically expanding what they can build autonomously.
• Vercel deployments: A direct build-to-ship pipeline that deploys your fullstack app to Vercel without leaving the platform.

Models now act as real agents

The architecture shift here is worth understanding. Models plan and execute autonomously using structured tool calls , create_file, edit_file, read_file , that reveal reasoning step-by-step. Models iterate, edit, and refine across multiple interactions, enabling complex builds within a single eval. With the fullstack update, that tool surface now includes bash execution and web search on top of file operations.

This is what separates Code Arena from a simple prompt-to-code tool. It emphasizes agentic behavior, allowing models to plan, scaffold, iterate, and refine code within controlled environments that replicate actual development workflows. Instead of checking whether code merely compiles, Code Arena examines how models reason through tasks, manage files, react to feedback, and construct functional web apps step by step.

Every action is also logged and traceable. Every model action , file creation, edit, or execution , is logged and versioned. Snapshots are stored in Cloudflare R2 and linked to Arena's database for transparent traceability.

What you can actually build now

The three categories of apps that unlock with this update are:

1. Auth-gated apps , Admin portals, dashboards, members-only storefronts. Anything that needs a sign-up and login flow backed by real user records in a database.
2. AI-powered apps , Chatbots, summarizers, or any product that calls an LLM API. You bring the key; the agent wires it up.
3. Persistent apps , Education platforms, habit trackers, progress-based tools where a user logs in over days or weeks and picks up where they left off.

These aren't toy demos. Production requirements are messy. Users don't ask for "a React component with proper prop typing." They ask for "a dashboard that shows user activity from the last 30 days with the ability to export to CSV." The frontend connects to the backend via real API calls. The backend performs actual database operations against PostgreSQL. Data persists between sessions. This is the same architecture developers use in production.

The dual purpose: build tool and benchmark

Code Arena occupies an unusual position: it's simultaneously a product you use to build software and a benchmark that evaluates which models are best at building software. Every build you run generates a vote that feeds the leaderboard. Arena's WebDev leaderboard has 328,594 votes across 81 models , and the fullstack votes will power a dedicated Fullstack leaderboard that doesn't exist yet.

Some models excel at database schema design but struggle with auth flows. Others generate solid APIs but don't wire them correctly to the frontend. Full-stack evaluation reveals these patterns. That's signal you can't get from HumanEval or SWE-bench.

For AI labs, this is also valuable training data. Each Code Arena evaluation is a reproducible experiment that captures the full trajectory of AI-assisted development, from ideas to generation to human judgment. Models interpret the request and decide which actions to take using structured tool calls , this agentic planning mirrors real developer workflow.

Industry context

Arena incorporated as an independent company in April 2025 and completed a $150 million Series A funding round in January 2026 at a $1.7 billion valuation. The fullstack launch is a product bet that the evaluation platform and the development platform can be the same thing , that the act of building software is itself the benchmark.

This puts Code Arena in direct competition with tools like Bolt.new and Replit, which also offer AI-assisted fullstack development in the browser. The differentiator Arena is betting on is the leaderboard feedback loop: every session contributes to a public, human-preference ranking of which models actually ship working software. At the beginning of the year, fewer than 3% of deployments on Vercel were triggered by coding agents; now agents account for more than half of all commits. The infrastructure moment is right.

Getting started

The platform is free , full access to all public arenas, leaderboards, and direct model testing, with no account required for basic use. AI Evaluations is the commercial service for structured model evaluations. You can try the Fullstack Code Arena now at arena.ai/code, and the team published a full walkthrough on their blog alongside a video demo on YouTube.

The clearest signal that this is more than a feature update: Arena is explicitly calling it a shift from "an experimental prototyping environment to a daily-use tool where real work gets done." That's a different product category, and the fullstack leaderboard coming next will tell us whether the models are actually ready for it.