Arena has introduced a new ranking system that incorporates factuality alongside human preference to assess model responses. This system allows users to view model rankings based on a weighted combination of these two signals, enhancing the evaluation process. To support this, over 2 million claims from large language models (LLMs) have been labeled, with notable changes in model positions since the implementation, such as GPT-5.5 moving up to the #7 spot and Claude Fable 5 dropping to #2. The methodology extends the Bradley-Terry objective, ensuring that models are not penalized for abstaining from making claims during factuality assessments.

Meta: Meta is the company behind open-source models evaluated in the Arena. Its overall lab ranking dropped when switching from pure human preference to the combined factuality metric.
Arena: Arena is the platform operating the Chatbot Arena, including Text Arena and Search Arena, that ranks large language models based on user-submitted comparisons. It has introduced factuality as a complementary signal to human preference, enabling weighted rankings that audit responses for verifiable claims extracted from real conversations. This update affects how models are positioned when users toggle the factuality view.
Claude: Claude refers to the model family developed by Anthropic and featured in Arena evaluations. The specific Claude Fable 5 variant experienced a minor downward shift in the Text Arena rankings under the new factuality-weighted system.
Xiaomi: Xiaomi produces open models tracked in Arena leaderboards. Among open-model providers, it achieved the strongest improvement in lab ranking under the new factuality-inclusive system.
GPT-5.5: GPT-5.5 is an OpenAI model included in Arena’s model comparisons. It recorded the largest upward movement among evaluated models when factuality was incorporated into the ranking calculation.
Anthropic: Anthropic develops the Claude model series and maintains top lab standing in Arena evaluations. The company preserved the number-one lab position even after factuality was added to the ranking methodology.
Muse Spark: Muse Spark is a model participating in Arena’s Text Arena rankings. It saw the steepest decline in position once the factuality component was applied to the composite scoring.

`json
{
“Model Evaluation”: “Arena evaluates model responses by sampling interactions, extracting verifiable claims, and assessing average correctness across models.”,
“Abstention Policy”: “Models are not penalized for choosing not to make claims during factuality evaluations.”,
“Ranking Methodology”: “The ranking system combines factuality with human preference using a composite score.”
}
`