A perfect score on the hardest math contest in the world
The International Mathematical Olympiad is a six-problem, nine-hour proof marathon that humbles professional mathematicians. The gold-medal threshold has historically sat around 29 out of 42 points. This year, Claude Opus 5 scored 42 out of 42 — without writing a single line of Python, calling an external solver, or using any agent scaffolding.
The result was announced by researchers tracking AI benchmark performance shortly after the IMO 2026 results were finalized. Anthropic generated four independent solutions to each of the six problems, had all 24 checked by an automated judge, and then had a human grader score a selected solution for each problem. All six received full marks.
How the run worked
Anthropic prompted the model to produce a rigorous, self-contained proof for each problem with a 256,000-token output limit and adaptive thinking set to maximum. One solution ran out of output space and was resampled at lower thinking effort; that was the only intervention. No code execution, no symbolic math libraries, no tool calls — just long-form mathematical reasoning in natural language.
The IMO 2026 leaderboard at BenchLM.ai shows Opus 5 at the top with 42/42. Multiple competing frontier models also attempted the problems: Claude Fable 5, Google's Sol, and others all achieved perfect scores as well, suggesting the 2026 IMO problems have effectively been solved as a benchmark category. The bar has moved.
What this actually means
The first honest answer is: this benchmark is now saturated. IMO problems, once considered a credible ceiling test for AI mathematical reasoning, no longer distinguish between the leading models. The community will need harder targets — ongoing research competitions, unsolved conjectures, or novel proof-generation tasks.
The second answer is more practical. Opus 5 was released on July 24, 2026, positioned as Anthropic's everyday enterprise model — a step below the company's flagship Fable 5 in cost ($5 per million input tokens vs. Fable's higher tier) but designed for extended agentic work, with a 1M-token context window and up to 128K output tokens. That a model in this price bracket can now solve every IMO problem should change how developers think about deploying mathematical reasoning tasks. Problems that previously required the most expensive tier may not anymore.
The third answer is the one that sits with you: solving all six IMO problems on the first pass, with multiple independent correct proofs per problem, places current AI performance "roughly five standard deviations above average human capability" on these tasks, according to MindStudio's benchmark analysis. Mathematicians who spend careers working on olympiad-style problems would score somewhere around the gold threshold. The model cleared it by 13 points.
What comes next
The IMO benchmark joins a growing list of once-formidable tests that AI systems have fully saturated: coding competitions, bar exams, professional licensing tests. Each saturation point reshapes the conversation about what to measure next. For mathematical reasoning, the frontier is now graduate-level open problems and real-time research collaboration — domains where correctness is harder to verify and the search space is genuinely open.
For now, the headline stands: a model you can access today via API solved every problem the world's best high-school mathematicians attempted this year. Without a calculator.