Vibe Code Bench: Can AI Build a Web App From Scratch?

Vibe Code Bench does not test isolated code snippets. It delivers a web application description and asks the model to build the app from scratch, with interface, logic and tests. It is an end to end software engineering test, not autocomplete.

In this article you will see:

• What Vibe Code Bench asks the model to do.
• How the evaluation works (steps, limits, pass criteria).
• The official ranking from November 5, 2026 and what the numbers show.
• Why this benchmark matters for people who use code agents.

What is Vibe Code Bench

Vibe Code Bench is a benchmark that evaluates whether a model can build a web application from a natural language description. Each task asks for a functional app, with interface, business logic and tests. The model works alone, without internet access, and delivers the complete source code.

The idea is to measure building ability, not correction. The model needs to make architecture decisions, implement the interface, write the logic and ensure tests pass. This is closer to real developer work than fixing a bug in an existing repository.

This differentiates Vibe Code Bench from benchmarks like CursorBench, which measures the ability to edit code in an editor. Here the model builds from scratch. It also differs from SWE-bench Pro, which starts from an existing repository and asks for a patch.

How the evaluation works

Vals AI evaluates models on 125 public tasks from Vibe Code Bench. Each task provides a web application description. The agent receives a bash terminal, an offline sandbox and the benchmark's standard prompt. It can inspect the delivered files, but it cannot use the internet, consult package registries or use binary analysis tools.

The submission contains source code files and a build script. Vals compiles and runs the tests in a new container. The agent has up to 1,000 steps, a 6 hour wall clock limit and a 180 second timeout per action. Tool output is truncated to 10,000 characters at the edges.

An app passes if the interface renders, the logic works and the hidden tests pass. There is no partial credit for explanation: either the app works, or it does not.

The official ranking from November 5, 2026

Vals tested 125 models on Vibe Code Bench. Claude Sonnet 5.5, running on OpenHands, leads with 92.39%, followed by Claude Sonnet 5 with 89.50% and Claude Sonnet 4.6 with 85.20%. Most models fall below 70%.

The gap between first and second is almost 3 percentage points. This shows that building web apps from scratch is a task that frontier models already handle reasonably well, but with room for error.

Vals also compares models on cost and latency per task. Claude Sonnet 5.5 on OpenHands is more expensive per attempt, but it is the only one that passes 92%.

How to test today

For anyone who wants to try it out, Vals makes the methodology and results available on the official leaderboard. You can run the same models on your own project and compare the results with the published ranking.

Vals also publishes the evaluation code and the prompts used, which makes it possible to reproduce the tests on other models and environments. This is useful for teams that want to validate whether a new model meets their needs before adopting it in production.

Why this benchmark matters

Vibe Code Bench measures a capability that patch benchmarks do not capture: the ability to build software from scratch from a description. This is useful in three real world scenarios.

First, rapid prototyping: when you need a working app in hours, not days. Second, validating code agents that need to deliver complete features, not just fixes. Third, comparing models to choose which one to use on build projects.

The 92.39% result is also a reminder of the limits. Even the best model is wrong in almost 8% of cases. For people who use code agents day to day, this reinforces the need for human review and testing, as discussed in the analysis of FeatureLoop in real repositories.