Anthropic’s Claude 5 has achieved 87.3% on the GPQA Diamond benchmark ” — reportedly the first time any AI system has exceeded 85% on what’s considered one of the hardest reasoning tests available. The benchmark results from March 3 represent an 8.1 percentage point jump over the previous record, equivalent to roughly four years of prior progress compressed into a single model update.

What Makes This Different

GPQA Diamond doesn’t test pattern matching. Each question requires genuine scientific reasoning in biology, chemistry, physics, or mathematics, with questions that take PhDs 2-3 hours to answer correctly. Questions include plausible but incorrect “distractor” answers and cannot be solved through memorization.

Extended Thinking Made the Difference

Standard Claude 5 mode scored 72.1%. Extended Thinking mode ” — the paid reasoning feature ” — jumped to 87.3%. That 15-point improvement came from inference-time reasoning optimization, not additional training data or larger model size.

Anthropic’s Chief Scientist commented: “This breakthrough confirms our thesis: reasoning is learnable, and scale alone was never the path forward.”

The Catch

Extended Thinking requires 40-50x more tokens, significantly increasing costs. And with human expert agreement on GPQA at only 87.9%, we may be approaching a ceiling.

OpenAI quickly released benchmark results from an unreleased GPT-5.1 model claiming 85.7%, while Google committed to GPQA Diamond focus in upcoming Gemini updates.

Source: Anthropic, March 3, 2026