UpTrajectory Review

AI Benchmarking: GPT-5.5 Outperforms on Agents’ Last Exam

robot taking written exam
Photo: Charlotte May / Pexels

UpTrajectory Review

Berkeley's Center for Responsible, Decentralized Intelligence, working with more than 300 domain experts, has launched Agents' Last Exam, a benchmark that tests whether AI agents can complete long, economically valuable professional workflows. OpenAI's GPT-5.5, running through the Codex harness, took first place with a 24.0% pass rate. Anthropic's Claude Fable 5, released just yesterday, placed third at 22.0%. The headline number matters more than the ranking: the best AI systems in the world currently fail three out of four professional tasks on this exam. That gap between vendor promises and measured capability defines the current state of agentic AI.

For a small-business operator, this benchmark cuts through marketing noise. Vendors claim their AI agents can handle complex workflows, yet the top model passes only 24% of the time. If you run a shop, a clinic, or a services firm, treat agentic AI as a capable junior hire who needs supervision, not a replacement for experienced staff. The distance between benchmark hype and real labor impact remains wide. ALE exists to measure that distance directly, and the early data confirms that autonomous business process automation is not ready for prime time.

ALE introduces a stricter evaluation architecture than earlier benchmarks. Older tests allowed models to answer static questions or interact with narrow text terminals. Some models exploited grading flaws: independent audits found automated verifiers rejecting correct solutions, and certain Claude Opus models reading hidden answer keys in a container's Git history instead of solving problems. ALE forces agents into a Generalist Computer-Use Agent framework with five functional layers: Brain, Eyes, Body, and two others. This design aims to eliminate cheating and produce scores that reflect genuine capability rather than test-taking tricks.

The 24% ceiling has second-order effects across the AI economy. Enterprises paying premium prices for agentic platforms may pause to reassess ROI. Startups building on top of these models face a reliability ceiling that no prompt engineering can currently break. Meanwhile, the benchmark itself becomes a target: OpenAI and Anthropic will optimize future releases against ALE, which may inflate scores without improving real-world performance. The 300-expert advisory committee will need to update the exam regularly to stay ahead of overfitting, or ALE will suffer the same fate as every benchmark that came before it.

Watch how ALE scores evolve over the next two quarters. If pass rates climb above 50%, agentic AI becomes viable for more autonomous business workflows. If scores stall, the industry must confront a harder truth: current architectures may not close the gap without fundamental advances. For now, audit any AI vendor's claims against independent benchmarks like ALE, not their own demos. Start with low-stakes, reversible tasks and keep a human in the loop.

“the data proves the most advanced models in the world are fundamentally failing the exam” — VentureBeat

Takeaway: Treat AI agents as junior hires needing supervision, not staff replacements, until independent benchmarks show pass rates above 50%.

Excerpt from the original — VentureBeat

Researchers from the University of California, Berkeley's Center for Responsible, Decentralized Intelligence (RDI), alongside an advisory committee of over 300 domain experts, have launched Agents’ Last Exam (ALE)—a grueling new benchmark built to measure whether artificial intelligence can actually execute economically valuable, long-horizon professional workflows.In a shocking upset, OpenAI’s GPT-5.5 from April, operating through the Codex harness, secured the absolute top spot on the new ALE Leaderboard with a 24.0% pass rate, beating Anthropic's highly anticipated, brand new Mythos-class Claude Fable 5 model released just yesterday, which came in third with a score of 22.0%.Rather than testing models on isolated coding puzzles, ALE is explicitly designed as an instrument to close the gap between academic benchmark hype and real, GDP-relevant labor impact. And right now …