Google is trying to measure whether AI agents can do the work of practicing engineers, not just answer textbook questions. On November 3, 2026 the Google AI & Infrastructure team announced Physical-Engineering's Last Exam (PLE), a benchmark and open competition built with the expert-network company Mercor. Its first results are sobering: every frontier model tested scored below 50% on the private test set.
PLE contains 85 long-horizon tasks spread across four disciplines: 30 mechanical, 27 electrical, 18 civil and 10 structural. Instead of a single prompt, an agent is dropped into a containerized project folder, built on Mercor's open-source Archipelago environment, that resembles what an engineer receives on the job. Tasks hold 61 files on average, 5,170 in total, including PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models and document revisions. The agent has to find the governing codes and physical limits, run the calculations and produce deliverables a licensed reviewer could sign off on, for example sizing a switchboard and its cable schedule or laying out a campus with grading and utilities.
Scoring is deliberately strict. Agents earn points only if a design first passes every hard safety and feasibility check, after which code compliance and cost efficiency are graded. Automated scripts check quantitative outputs and an LLM judge grades open-ended work against expert rubrics. Each task runs without internet access for up to six hours, and every model and harness pair is run three times.
Forty-four of the tasks were written by practicing engineers from Mercor's network, who built the folders, prompts and rubrics themselves. The other 41 were co-authored: Google used an adversarial loop in which one agent drafted a task and another tried to break it, and engineers then validated each task and raised its difficulty. Google and Mercor released 29 tasks publicly with grading tools and an evaluation container on GitHub; the remaining 56 form a hidden test set for the competition at ple-bench.com, where 13 Google-sponsored models can be run free. A paper describing the benchmark is on arXiv.
Fourteen frontier models were evaluated with seven public agent harnesses, among them Antigravity, Claude Code, Codex, Hermes, OpenCode, OpenHands and Pi. On their native harnesses, Claude Opus 5.5 in Claude Code led at 48.3%, ahead of GPT 6 Astra in Codex at 45.8% and Gemini 4 Argon at 40.6%, according to Mercor. Google reports that the strongest configuration satisfied 91% of individual constraints on co-authored tasks, yet only 41.3% of its deliverables cleared every hard check. Mercor adds that the top runs averaged about 27 million tokens and 63 minutes per task.
The headline finding concerns where performance comes from. Scores across models spread by up to 50 points, while the same model on different harnesses moved by a median of about 3 points, although harness choice changed costs sharply. The public leaderboard shows Opus 5.5 costing $42.32 per task with Antigravity against $12.87 with OpenCode.
Some caution is warranted. PLE is a vendor-built benchmark, its tasks are new and have not yet been widely reproduced, and the leaderboard lists Claude Opus 5.5 as the rubric judge, the same model family that tops the table. Small public samples also make overfitting easier once the competition starts. Still, the benchmark targets a real gap, since most agent evaluations stop at software, and it gives developers a concrete measure of how far agents remain from engineering work that has to be safe as well as correct.