garak vs LiveBench

garak is a command-line scanner for probing language models and systems that use them for security weaknesses. Its checks cover issues such as hallucination, data leakage, prompt injection, misinformation, toxicity generation, and jailbreaks. Probes try to elicit failures, while detectors inspect responses for patterns such as unsafe content or a failure to refuse. A standard scan can run without supervision and produce a report, with scan settings available for customization. Integrations include OpenAI, Hugging Face, Cohere, Replicate, AWS Bedrock, LiteLLM, REST-accessible models, and custom Python integrations. Hugging Face models can run locally or be accessed through an inference API or private endpoint. Output includes screen results, a detailed report log, a hit log, and a debug log. garak assesses text output only. It is developed in Linux and OSX and can be installed with pip or from source; listed platforms include Linux, macOS, Windows, and self-hosted use. It is free, with a free plan and no free trial. Listed capabilities include custom metrics, LLM-as-a-judge, and safety evaluations.

LiveBench is an open-source benchmark for evaluating large language models, designed to limit test-set contamination and use objective scoring. Its current scope is 23 tasks across Reasoning, Coding, Agentic Coding, Mathematics, Data Analysis, Language, and Instruction Following, with a refresh every six months. The leaderboard presents overall and category scores, and model rows expand to show subtasks. Insight views cover quality versus cost, ranked cost, and category profiles; cost per successful task is calculated as (Σ cost ÷ Σ questions ÷ score) × 100 for the selected scope. Questions have verifiable ground-truth answers and can be scored automatically without an LLM judge. The public project provides benchmark code and links to data on Hugging Face, as well as a Python command-line evaluation pipeline. It supports OpenAI-compatible endpoints and implemented inference for Anthropic, Cohere, Mistral, Together, and Google models. Local inference is unmaintained. Agentic coding evaluations require Docker, and task images may take up to 150 GB to store.