AI Agent Scientific Task Measurement: The strongest agent was only able to complete 20.6%.

Watchtower Beating AI News Flash: Chen Tianqiao's new AI project, Apodex, has released the Scientific Agent Evaluation FrontierChallenge, testing whether AI can complete a task from start to finish across 97 real-world scientific workflows. 12 cutting-edge models participated in the test, with the highest score reaching only 20.6%. GPT-5.6 Sol + Codex and Grok 4.6 + Claude Code tied for first place, each successfully completing all 20 tasks.

The test also found that Agents often claimed to have completed tasks when they had not. In a test of 10 groups of models using Claude Code as the Agent framework, out of 849 failed tasks, 75.5% were falsely reported as completed. On the other hand, all evaluated systems achieved an average score of 94.9 in tasks related to electrochemistry and environmental science, but none completed all tasks.

This evaluation specifically distinguishes between "getting most right" and "truly delivering completion." The ranking comparison is based on the model + Agent operating framework, not just a comparison of bare models.

The official statement also emphasizes that each system has only one run per task, and slight score differences should not be directly interpreted as stable model rankings.