July 7, 2026

Artificial Analysis launches Harvey LAB-AA to benchmark legal models

Artificial Analysis has launched Harvey LAB-AA, a new benchmark designed to evaluate language models on their performance in real-world legal tasks across 24 practice areas, including corporate transactions and litigation. The benchmark consists of a set of 120 legal tasks that the models must complete, graded against binary criteria that assess their ability to deliver complete legal outputs. Leading the evaluation is Claude Fable 5, which achieved a 14.2% all-pass rate, signifying that it fully met the requirements of only a fraction of the tasks. This statistic highlights a key finding of the benchmark: while many models can pass individual criteria, very few can satisfy all requirements, indicating that frontier legal work remains a challenging area for AI models.

Zai: Zai is the organization behind the GLM series of language models. Its GLM-5.2 model tied for second place in the Harvey LAB-AA benchmark results.
Harvey: Harvey develops AI tools tailored for legal professionals and created the original LAB benchmark for evaluating language models on legal work. Artificial Analysis built Harvey LAB-AA as an independent version of this benchmark to test models across multiple practice areas.
GLM-5.2: GLM-5.2 is a language model developed by Zai. It tied for second place in the Harvey LAB-AA benchmark.
GPT-5.5: GPT-5.5 is a language model developed by OpenAI. It was evaluated in the Harvey LAB-AA benchmark.
Anthropic: Anthropic is an AI research company that creates the Claude family of language models. Its Claude Fable 5 model led the legal benchmark results, with other Claude variants also participating in the evaluation.
MiniMax-M3: MiniMax-M3 is a language model that participated in the Harvey LAB-AA evaluation. It placed among the leading models in the benchmark results.
Claude Fable 5: Claude Fable 5 is a frontier language model developed by Anthropic. It achieved the top position in the Harvey LAB-AA benchmark among all evaluated models.
Claude Opus 4.8: Claude Opus 4.8 is a language model from the Claude family developed by Anthropic. It tied for second place in the Harvey LAB-AA benchmark.
Claude Sonnet 5: Claude Sonnet 5 is a language model from the Claude family developed by Anthropic. It participated in the Harvey LAB-AA benchmark and ranked among the top performers.
Gemini 3.1 Flash: Gemini 3.1 Flash is a language model developed by Google. It was evaluated in the Harvey LAB-AA benchmark.
DeepSeek V4 Flash: DeepSeek V4 Flash is a language model from the DeepSeek series. It was evaluated in the Harvey LAB-AA benchmark.
Artificial Analysis: Artificial Analysis is an organization focused on independent evaluations and benchmarking of artificial intelligence models. It recently launched Harvey LAB-AA as its reimplementation of an existing legal agent benchmark, using its own agent harness and simplified prompts to assess model performance on legal tasks.

Industry Focus: The benchmark targets real-world legal practice areas including corporate transactions, litigation, tax, and bankruptcy.
Benchmarking Approach: Harvey LAB-AA uses a private set of legal tasks and binary rubric criteria to evaluate models on complete professional deliverables rather than partial criteria.
Model Performance Patterns: Frontier models in the evaluation tend to engage in extended agentic loops and generate substantial output tokens per task.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like