Anthropic has announced that its new model, Claude Fable 5, achieves superior results on the FrontierMath benchmark compared to OpenAI’s GPT-5.5. According to Epoch AI, Fable 5 scores 87 percent accuracy on tiers one through three and reaches 88 percent on the most difficult tier four. This performance represents a dramatic improvement over the company’s earlier Opus 4.5 model, which scored below 10 percent on that same tier in early 2026. Meanwhile, OpenAI’s GPT-5.5 manages approximately 75 percent on tier four, leaving a significant gap despite reports that GPT-5.6 is currently in development. All models were evaluated using Epoch AI’s standard scaffold with maximum reasoning effort, ensuring a consistent comparison across the industry leaders.
This development matters because FrontierMath is widely regarded as one of the most rigorous tests for artificial intelligence mathematical reasoning. The rapid progression from near-failure to near-perfect scores on tier four within a short timeframe suggests that current models are approaching the limits of what can be solved computationally. Real-world applications are already validating these benchmark gains, with both OpenAI models and Claude Mythos recently solving a longstanding Erdős problem. As these systems handle increasingly complex calculations, the implications for scientific research and automated problem-solving become more tangible for businesses and researchers alike.
- Claude Fable 5 scores 88 percent on FrontierMath tier four, significantly outperforming GPT-5.5’s 75 percent.
- Anthropic’s math capabilities have improved drastically since Opus 4.5 scored under 10 percent on the hardest tier in early 2026.
- Recent real-world successes, such as solving an Erdős problem, confirm that benchmark improvements translate to practical utility.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
