Physical-Engineering’s Last Exam (PLE) is a new benchmark and open competition from Google’s AI & Infrastructure team, developed in collaboration with Mercor, to test a frontier model’s ability to complete end-to-end tasks in mechanical, electrical, civil, and structural engineering. The release includes 29 public tasks, their grading tools, and an open competition leaderboard at ple-bench.com, plus a technical paper on arXiv and a GitHub repository with grading scripts and an evaluation container. PLE fills a gap the team found: no existing benchmark tests whether frontier models can do physical-engineering work end to end, even though physical engineers design skyscrapers, power grids, and computer chips. The benchmark is framed alongside Humanity’s Last Exam for general knowledge and Agents’ Last Exam for digital professional work.

Existing academic benchmarks fall short for physical engineering, according to the team. Frontier language models score above 90% on many academic benchmarks, but those mostly test textbook-style questions, single tool calls, or small self-contained design problems. They do not show whether an agent can produce safe, code-compliant, and cost-effective designs under a licensed engineer’s review and control. PLE addresses this by giving the agent a containerized project folder, built on Mercor’s Archipelago environment, that mirrors what a practicing engineer receives. Each task holds an average of 61 files (5,170 total across the benchmark), including PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models, and document revisions. The agent must reason through these files, find the governing building codes and physical limits, run calculations, and generate production-ready deliverables that a licensed reviewer would approve.

The benchmark is 85 long-horizon tasks across four disciplines: mechanical (30 tasks), electrical (27), civil (18), and structural (10). PLE was built with experts from Mercor and licensed practicing engineers using two methods. In 44 expert-authored tasks, practicing engineers created the project folders, prompts, and grading rubrics from real professional engineering tasks. In 41 co-authored tasks, engineers designed the initial task and an agent pipeline assembled the task’s environment and grader from licensable engineering standards; engineers then validated and increased the difficulty of every task assembled by the agent.

Google evaluated 14 frontier models on PLE with seven public agent harnesses: Antigravity, Claude Code, Codex, Hermes, OpenCode, OpenHands, and Pi. Because PLE evaluates both code compliance and cost efficiency, agents only earn points if their design passes every hard safety check first. Baseline results show a wide gap between current models and real-world engineering requirements. All frontier models score below 50% on the private test set, with the highest-performing model on its native harness reaching a mean score of 48.3%. The top-scoring agent configuration satisfied 91% of individual constraints on co-authored tasks, but only 41.3% of its deliverables passed every hard safety and feasibility check required. Choice of model drives performance far more than the harness: up to a 50-point spread across models, versus a median 3-point spread across harnesses for the same model, though harnesses significantly affect run cost.

The team frames the work as a practical step toward helpful tools for physical engineers, especially as infrastructure demand outpaces financing. Researchers and developers can download the 29 public tasks, grading scripts, and evaluation container from GitHub, read the full technical breakdown in the PLE paper on arXiv, and submit their own agent setup for evaluation on the private test set at ple-bench.com. Google is sponsoring 13 models running free on the servers, with prize details coming soon.