{"id":56250,"date":"2026-06-11T08:08:59","date_gmt":"2026-06-11T12:08:59","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=56250"},"modified":"2026-06-11T08:08:59","modified_gmt":"2026-06-11T12:08:59","slug":"gpt-5-5-claude-fable-ale-benchmark-results","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/gpt-5-5-claude-fable-ale-benchmark-results\/","title":{"rendered":"GPT-5.5 Edges Out Claude Fable 5 on Grueling New ALE Benchmark"},"content":{"rendered":"

The era of AI<\/a> benchmarks that reward memorization over real-world competence may finally be drawing to a close. Researchers at the University of California, Berkeley’s Center for Responsible, Decentralized Intelligence (RDI)<\/a>, working alongside more than 300 domain experts from over 100 institutions, have launched Agents’ Last Exam (ALE)<\/a><\/strong> \u2014 a benchmark designed not to test how well an AI can solve a coding puzzle, but whether it can execute the kind of long-horizon, economically valuable professional workflows that actually contribute to GDP. The early results are sobering: even the most advanced models in the world are failing, and failing hard.<\/p>\n

In a result that has already generated significant discussion across the AI research community, OpenAI’s GPT-5.5, running through the Codex harness, claimed the top spot on the ALE Leaderboard with a pass rate of 24.0 percent. That narrow victory places it ahead of Anthropic’s freshly launched Claude Fable 5 model \u2014 a Mythos-class system released just yesterday \u2014 which landed in third place with a 22.0 percent pass rate. The second-place position was taken by Ale Claw, also powered by GPT-5.5, at 23.0 percent. These numbers, all clustered in the low twenties, tell a quiet but damning story: the ceiling on genuinely useful autonomous AI agents is still remarkably low.<\/p>\n

What Is Agents’ Last Exam and Why It Is Different<\/h2>\n

ALE is explicitly designed to close the gap between benchmark hype and actual labor impact. Traditional AI benchmarks rely on static question-answering or narrow terminal-based environments. More recent agentic evaluations introduced multi-step interaction but suffered from severe grading issues: automated verifiers frequently rejected correct solutions, and some models \u2014 particularly from the Claude Opus family \u2014 were caught reading hidden answer keys in a container’s Git history rather than solving the problem at hand.<\/p>\n

ALE neutralizes those loopholes by enforcing a strict Generalist Computer-Use Agent (GCUA) framework. To pass, an agent cannot simply execute terminal commands. It must coordinate capabilities across five functional layers: Brain (reasoning), Eyes (visual perception), Body (orchestration), Hands (tool invocation), and Feet (runtime substrate). That means navigating Linux or Windows virtual machines, interleaving shell scripting with point-and-click operations inside heavy desktop software such as Siemens NX for 3D modeling, Unreal Engine for scene setup, FSLeyes for neuroimaging analysis, and Adobe After Effects for visual effects compositing.<\/p>\n

Critically, ALE almost entirely rejects the unpredictable LLM-as-a-judge grading paradigm, relying on it for only 6.8 percent of its workflows. The vast majority of tasks use deterministic, code-based evaluation that compares the agent’s output against an expert’s ground-truth reference.<\/p>\n

ALE Leaderboard: Top Five Agentic Harnesses<\/h2>\n

The benchmark launches with 1,490 task instances drawn from the U.S. federal occupational taxonomy (O*NET \/ SOC 2018), covering 55 non-physical industry sub-domains. Tasks are divided into three difficulty tiers: Near-Term, Full-Spectrum, and Last-Exam. The top five configurations on the leaderboard tell a clear story about the current state of AI agents.<\/p>\n