LiveBench
A continuously updated evaluation benchmark for large language models that mitigates data contamination by regularly refreshing its questions. It is used for fair comparison of the overall capabilities of different models.
Timeline (last 90 days)
On LiveBench long-horizon agentic tasks, Opus and Astra both score above 80, while Argon scores only about 70.
LiveBench's agentic coding sub-item specifically examines a model's sustained performance in multi-step tool calls and file operations.
LiveBench's questions come from recent competitions, papers, and news, are updated monthly to prevent data contamination, and have objectively verifiable answers.
Academia is exploring dynamic benchmarks (such as HELM and LiveBench) and adversarial evaluation (Adversarial NLI) as alternative approaches.
Referencing the LiveBench benchmark, Claude Fable 5.1 and GPT-6 Astra have comparable capabilities.
New-generation evaluation frameworks such as LiveBench periodically replace questions to combat data contamination.
LiveBench is a dynamic benchmark that updates its test questions monthly.
Many classic benchmarks have become saturated (e.g., multiple models scoring above 95% on HellaSwag), and the industry is developing new benchmarks with greater discriminative power, such as GPQA, SWE-bench, and LiveBench.
All Facts (7)
On LiveBench long-horizon agentic tasks, Opus and Astra both score above 80, while Argon scores only about 70.
50%UnverifiedLiveBench's agentic coding sub-item specifically examines a model's sustained performance in multi-step tool calls and file operations.
50%UnverifiedLiveBench's questions come from recent competitions, papers, and news, are updated monthly to prevent data contamination, and have objectively verifiable answers.
50%UnverifiedAcademia is exploring dynamic benchmarks (such as HELM and LiveBench) and adversarial evaluation (Adversarial NLI) as alternative approaches.
50%UnverifiedNew-generation evaluation frameworks such as LiveBench periodically replace questions to combat data contamination.
50%UnverifiedLiveBench is a dynamic benchmark that updates its test questions monthly.
50%UnverifiedMany classic benchmarks have become saturated (e.g., multiple models scoring above 95% on HellaSwag), and the industry is developing new benchmarks with greater discriminative power, such as GPQA, SWE-bench, and LiveBench.
50%