Anthropic's Claude Fable 5 has achieved 88 percent accuracy on the toughest tier of FrontierMath, a benchmark built around research-level mathematics problems — and the result puts it clearly ahead of its biggest rival.
According to The Decoder, the score marks a dramatic leap from Anthropic's previous flagship, Opus 4.5, which sat below 10 percent on the same benchmark tier as recently as early 2026. That is an improvement of more than 78 percentage points within months.
OpenAI's GPT-5.5 reaches roughly 75 percent on the same tier, according to The Decoder, putting Claude Fable 5 about 13 points ahead. The Decoder also notes that the pace of improvement in AI math keeps accelerating — a trend that shows no sign of slowing.
Anthropic has introduced Fable 5 as its first so-called "Mythos class" model, a new category in the company's lineup, according to The Indian Panorama.
FrontierMath is designed to challenge AI with problems at the level of professional mathematicians, not standard textbook exercises. Reaching near-90 percent accuracy on its hardest section would have seemed far-fetched just a year ago.
The stakes go well beyond a leaderboard: mathematics is a stand-in for the kind of rigorous, step-by-step reasoning needed to verify scientific findings, write provably correct code, and solve hard engineering problems — making these benchmark leaps a signal that AI's practical capabilities are advancing faster than most expected.