All Benchmarks/Agents

Vibe Code Bench v1.1

Often citedHigher is better
What does it measure?
Made by Vals AI: whether an AI can build a web application from scratch, working only from a natural-language spec, and get it actually running.
Who evaluates it, and how?
For each non-technical spec of under a page, the AI builds an app, and an evaluator agent runs 20–60 automated tests per app by clicking through it. Product managers and engineers reviewed the specs and tests.
Reading the score
Scores run from 0 to 100%. Higher is better. Top models now exceed 90%, so it separates them less. The test set is not public.
Top score93.8%Claude Fable 5.2
Models tested86
Last updated2026.11.09
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Anthropic
    Claude Fable 5.2
    2026.11.09 · OpenHands
    93.8%
  2. 2
    Anthropic
    Claude Sonnet 5.5
    2026.09.28 · OpenHands
    92.4%
  3. 3
    Google
    Gemini 4 Argon
    2026.09.30 · OpenHands · high
    91.9%
  4. 4
    Anthropic
    Claude Haiku 5.5
    2026.10.07 · OpenHands
    90.4%
  5. 5
    Anthropic
    Claude Fable 5
    2026.06.09 · OpenHands
    90.4%
  6. 6
    Anthropic
    Claude Opus 5.5
    2026.09.22 · OpenHands
    90.3%
  7. 7
    Anthropic
    Claude Fable 5.1
    2026.09.01 · OpenHands
    90.3%
  8. 8
    OpenAI
    GPT-6 Astra
    2026.09.04 · OpenHands · max
    89.6%
  9. 9
    Google
    Gemini 4 Carbon
    2026.10.28 · OpenHands · high
    89.1%
  10. 10
    OpenAI
    GPT-6.1 Sol
    2026.09.29 · OpenHands · max
    88.9%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.

Where the scores come from

  • Run by the test’s team: scores the team that operates the test ran and published itself.