GPT-5.5 Surprises with Top Score in Agents' Last Exam Benchmark

Researchers from the University of California, Berkeley's Center for Responsible, Decentralized Intelligence (RDI) have launched Agents' Last Exam (ALE), a demanding new benchmark aimed at assessing whether AI can execute economically valuable, long-horizon professional tasks. In an unexpected twist, OpenAI’s GPT-5.5, operating through the Codex harness, secured the top spot on the ALE Leaderboard with a 24.0% pass rate. This beat Anthropic's highly anticipated Claude Fable 5 model, which came in third with a score of 22.0%.

The New Benchmark: ALE

Unlike traditional benchmarks that test models on isolated coding puzzles or narrow text-based environments, ALE is designed to bridge the gap between academic hype and real-world labor impact. The benchmark evaluates AI models across five functional layers: Brain (reasoning), Eyes (visual perception), Body (orchestration), Hands (tool invocation), and Feet (runtime substrate).

• Brain (Reasoning): The model's ability to reason through complex problems and make logical decisions.
• Eyes (Visual Perception): Visual recognition capabilities, essential for navigating graphical user interfaces (GUIs).
• Body (Orchestration): Coordinating tasks across multiple tools and environments.
• Hands (Tool Invocation): Executing specific commands or actions within software applications.
• Feet (Runtime Substrate): The underlying infrastructure that supports the model’s operations.

To pass ALE, an agent must navigate Linux or Windows virtual machines, interleaving shell scripting with point-and-click operations inside heavy desktop software. This comprehensive evaluation ensures that models are not just executing terminal commands but can handle a wide range of tasks in a realistic professional setting.

Key Takeaways

• GPT-5.5's Success: OpenAI’s GPT-5.5, released in April, demonstrated strong performance on ALE, achieving a 24.0% pass rate.
• Claude Fable 5's Performance: Anthropic’s Claude Fable 5, launched just the day before, scored 22.0%, placing third.
• Benchmark Rigor: ALE is designed to be more rigorous than previous benchmarks, focusing on real-world professional workflows rather than isolated tasks.
• Evaluation Architecture: The GCUA framework ensures that models are evaluated across multiple functional layers, making it harder for them to "cheat" or rely on shortcuts.

The success of GPT-5.5 on ALE highlights the model's versatility and robustness in handling complex, multi-step tasks. While Anthropic’s Claude Fable 5 is a powerful model, its performance suggests that there are still significant challenges in achieving true agentic capabilities in AI systems.

As the field continues to evolve, benchmarks like ALE will play a crucial role in driving progress and ensuring that AI models can deliver real-world value.