GPT-5.5 Outperforms Claude Fable 5 in New Agents’ Last Exam Benchmark

Researchers at the Center for Responsible Decentralized Intelligence at the University of California, Berkeley (RDI) have introduced Agents’ Last Exam (ALE), a new benchmark designed to evaluate AI agents' ability to perform economically significant professional tasks. Unlike traditional tests, ALE focuses on real-world workflows, such as creating 3D models in Siemens NX, configuring scenes in Unreal Engine, and working with neuroimaging in FSLeyes.

In ALE, models are assessed across five key parameters: logical reasoning, visual perception, task orchestration, tool usage, and runtime environment interaction. The benchmark covers 55 industries and includes 1,490 tasks based on real-world professional scenarios. ALE emphasizes deterministic evaluation: instead of using LLMs as judges, it employs code checks to compare AI-generated artifacts with reference solutions.

The test results were unexpected: OpenAI’s GPT-5.5, operating via the Codex framework, claimed the top spot with a 24% success rate, outperforming Anthropic’s recently released Claude Fable 5, which scored 22%. However, even leading models performed poorly on the most complex tasks—most models, including Claude Opus 4.8 and Gemini CLI, failed entirely. This underscores the existing gap between academic benchmarks and real-world AI performance in professional settings.

To prevent data leaks and ensure objective evaluations, ALE uses a unique task rotation system: only 10% of the dataset is publicly available, while the remaining tasks are stored privately and updated regularly. This approach prevents benchmark contamination and ensures reliable results for corporate clients.