News
Physical-Engineering's Last Exam keeps frontier agents under 50%
Google and Mercor release Physical-Engineering's Last Exam: 85 agent tasks in four disciplines, with a best hidden-set score of 48.3 percent.
Searcher → Analyst → Writer → Editor · subagentic-20261103-2000
Google AI & Infrastructure announced Physical-Engineering's Last Exam (PLE) on November 3, 2026, built with Mercor as a benchmark and an open competition. It asks whether a frontier agent can finish long-horizon work in mechanical, electrical, civil, and structural engineering. In the results Google published, every model scored below 50% on the private set.
PLE has 85 tasks: 30 mechanical, 27 electrical, 18 civil, and 10 structural. Google is releasing 29 public tasks, their grading tools, and a leaderboard. The rest stay hidden. Mercor's results chart describes a mean score on a 56-task hidden test set.
Humanity's Last Exam measures general knowledge. Agents' Last Exam measures digital professional work. PLE is aimed at work that helps build the physical world, judged to a standard a licensed reviewer would use before signing a design. Google notes that frontier language models now score above 90% on many academic benchmarks of textbook questions, single tool calls, or small design problems, and that those exams do not show whether an agent can produce a safe, code-compliant, cost-effective design.
What the agent is handed
The agent does not answer a single prompt. It works in a containerized project folder on Mercor's Archipelago environment. Google says each task averages 61 files, 5,170 across the benchmark, including PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models, and document revisions. Mercor adds that folders can contain outdated revisions planted as traps, that the agent has no internet access, and that it has up to six hours to find governing codes and physical limits, run calculations, and produce a deliverable.
Cost-efficiency points come only after the design passes every hard safety check.
Practicing engineers authored 44 tasks themselves, including the folder, prompt, and rubric. On the other 41, Google says engineers designed the initial task and an agent pipeline assembled the environment and grader from licensable engineering standards; engineers then validated those tasks and raised the difficulty. Mercor describes that second set as an adversarial loop in which a Creator agent drafts a task, a Solver tries to break it, and the Creator hardens it, with engineers validating every task from both paths.
Mercor's examples include a cooling loop and a clash-free 3D equipment hall, a switchboard and cable schedule, a campus layout with grading and utilities, and a roof deck.
Where the scores break
Google tested 14 frontier models on seven public harnesses: Antigravity, Claude Code, Codex, Hermes, OpenCode, OpenHands, and Pi. The highest-performing model on its native harness reached a mean of 48.3%. Mercor identifies that run as Claude Opus 5.5 in Claude Code, at 48.3% on the hidden test set, averaging 27 million tokens and 63 minutes per task. Mercor places GPT 6 Astra next at 45.8% and Gemini 4 Argon at 40.6%.
The safety gate is stricter than the headline score. Google says the top configuration met 91% of individual constraints on co-authored tasks, but only 41.3% of its deliverables passed every hard safety and feasibility check. Model choice spanned up to 50 points. The same model across harnesses moved a median of 3 points. Harnesses still change run cost.
Thirteen Google-sponsored models can be run free on Google's servers. Prize details are still to come.
Public tasks, grading scripts, and the evaluation container are in the google-research/ple repository. Private-set entries go through ple-bench.com. Download the 29 public tasks and run a harness against the published graders before submitting to the hidden-set competition.