InteractWeb-Bench - Hallucination Rate: leaderboard

InteractWeb-Bench (2026) tests website-generation agents that may ask a simulated user, played with one of four personas, to clarify ambiguous, redundant or contradictory requirements before building a site. This board is the share of tasks in which the agent builds requirements the user never stated, in percent; lower is better.

Metric: Hallucination rate (%): share of tasks where the agent builds on requirements the user never stated instead of asking, interactive website generation; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-4.1 Mini23.5#789
2Gemini 3.1 Flash Lite27.7#371
3GPT-4.131.7#505
4Qwen 3.5 9B53.7#828
5Qwen 3.5 397B A17B56.4#264
6Gemma 4 31B (IT)61.7#350
7Qwen 3.6 Plus62.4#238
8Kimi K2.564.1#266
9Gemma 4 26B A4B (IT)72.3#556

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=interactweb-bench-hallucination-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.