AI Agent Scientific Task Field Test: The Strongest Performer Only Completes 20.6% of Tasks Fully

Dongcha’s Beating AI News reports that Apodex, a new AI project under Chen Tianqiao, has released FrontierChallenge, a scientific agent benchmark that tests whether AI can complete tasks end-to-end using 97 real-world scientific workflows. Twelve leading models participated in the test, with the best performance reaching only 20.6%. GPT-5.6 Sol + Codex and Grok 4.6 + Claude Code tied for first place, each fully completing 20 tasks. The test also found that agents often claim to have finished tasks even when they are incomplete. Among 10 model groups using Claude Code as the agent framework, 75.5% of 849 failed tasks were falsely reported as completed. On the other hand, all participating systems scored an average of 94.9 on electrochemistry and environmental science tasks, but none fully passed any task. This benchmark explicitly separates "mostly correct" results from "truly completed deliverables". The ranking compares models plus agent runtime frameworks, not bare models alone. The official also reminded that each system is allowed only one run per task, so small score gaps should not be directly interpreted as a stable model ranking.