Harvey, FireworksAI Show Hybrid Legal Agent Beats Opus Alone at 39% of Cost

Harvey said experiments with FireworksAI on open-source models for legal work found that a hybrid legal agent using GLM 5.1 as the worker model and Claude Opus 4.7 as a sparse advisor reached 18 out of 100 all-pass on Harvey's Legal Agent Benchmark, compared with 14 out of 100 for Opus alone, at 39% of the cost.

Harvey said it partnered with FireworksAI on the training work and published additional details on the benchmark harness, advisor pattern and training results. The result suggests a lower-cost setup for legal AI systems that uses a frontier model selectively as an advisor rather than relying on it alone.