AI Agent Performance Benchmarks: New Data on Fable 5 and Remote Labor

The latest testing from the Center for AI Safety (CAIS) indicates that specialized AI agents are reaching an automation rate of 16.1% on complex freelance tasks, according to the organization’s Remote Labor Index (RLI) released in late 2025. Anthropic’s Fable 5 model currently leads this benchmark, outperforming competitors like OpenAI’s GPT-5.5 and Anthropic’s own Opus 4.8 in executing professional-grade assignments ranging from graphic design to data analysis.

How AI Agents Perform on Real-World Freelance Tasks

The Remote Labor Index measures the frequency at which AI agents complete economically valuable projects at a quality level acceptable to professional clients. According to CAIS, the assessment requires models to handle multi-step tasks, such as designing a 3D engagement ring mockup or mapping a floor plan, using human-generated input files. Human evaluators then grade the outputs against professional standards. The 16.1% automation rate achieved by Fable 5 represents a significant increase compared to previous benchmarks, where the highest-performing models struggled to exceed 4% efficiency. Even when accounting for potential failures in incomplete test sets, CAIS researchers noted that the model maintained a baseline performance of 14.6%.

Comparison of Leading AI Models

Recent benchmarking data highlights a clear disparity in the task-completion capabilities of current frontier models. While all tested models showed improvements over earlier iterations, Fable 5 currently holds a distinct lead in the RLI rankings.

According to CAIS, the industry has seen agentic skills quadruple in less than eight months. When the RLI was first introduced, the top-performing models in the field reached an automation rate of only 2.5%.

Why Computer-Use Skills Remain a Barrier

Despite these gains, researchers identified persistent limitations in how AI agents interact with professional software. CAIS attempted to use an “LLM judge” to evaluate project deliverables automatically, but the experiment failed. The agency reported that evaluating work requires the same high-level, agentic computer-use skills—such as navigating professional applications and forming subjective quality judgments—that current models still lack. Because these models struggle to operate within the specific software environments used by professionals, they are not yet capable of replacing human freelancers in a one-to-one capacity.

The Future of AI in Remote Work

The acceleration of AI capabilities does not signal an immediate, broad-scale displacement of human labor. Integration remains a slow, multi-step process for most organizations due to security concerns and the necessity of human oversight for quality control. However, the rapid improvement in agentic benchmarks suggests that the current roadblocks may shrink as investment in specialized AI agents continues. CAIS observed that AI performance does not correlate linearly with the time a task takes a human; models often complete complex digital art or coding tasks in minutes, while struggling with simpler, real-time creative work like transcribing music.