Surprise upset: GPT-5.5 beats Claude Fable 5 on brutal new Agents’ Last Exam benchmark

Anthropic has achieved a significant competitive breakthrough with the launch of Claude Fable 5, its Mythos-class large language model, which demonstrates decisive performance advantages over OpenAI's GPT-5.5 across multiple standardized benchmarks. On Humanity's Last Exam—a benchmark designed to evaluate frontier AI capabilities—Fable 5 achieved a score of 91 out of 100 (or 53% on normalized leaderboard scoring), substantially outperforming GPT-5.5's score of 62 out of 100. This performance delta positions Fable 5 marginally below the range of actual human engineers, indicating capability progression toward expert-level problem-solving in complex domains.

The new model's advantages extend across specialized evaluation criteria relevant to institutional investors deploying agentic AI systems. On the Legal Agent Benchmark, Fable 5 significantly outpaced GPT-5.5, which achieved only 2.1% accuracy in standalone testing. Independent researchers including Matthew Pines, conducting frontier physics research evaluation, reported Fable 5 achieving superior performance across multiple scientific domains. The broader performance gap is reflected in Fable 5's #1 ranking on the Artificial Analysis Intelligence leaderboard, where it maintains more than a 7-percentage-point margin over the next-best model, Claude Opus 4.8.

Pricing represents a critical variable for institutional adoption: Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, positioning it at a meaningful cost premium relative to competitive models. However, multiple independent developers conducting hands-on testing in production environments report that Fable 5's capability improvements justify the cost differential. Real-world testing on actual production SaaS applications and complex coding workflows indicates that the benchmark performance gap translates to meaningful improvements in model reliability, reasoning depth, and task completion rates—with one evaluator characterizing GPT-5.5 as feeling like a "toy" by direct comparison.

The launch establishes a clear capability bifurcation at the frontier tier of the LLM market. For institutional investors and family offices evaluating AI infrastructure investments or deployment strategies, Fable 5's performance profile suggests Anthropic has advanced meaningful distance ahead of OpenAI in specific high-value application domains, particularly agentic reasoning, legal analysis, and technical coding tasks. The practical impact of this performance gap—rather than abstract benchmark differences—will likely drive platform selection decisions for mission-critical applications where accuracy and reasoning capability directly impact operational value.