Will an Anthropic Claude model score at least 45% on Humanity’s Last Exam?

Prediction markets put the probability at 14%: Will an Anthropic Claude model score at least 45% on Humanity’s Last Exam. Currently, markets see this as unlikely (14% YES). Anthropic is investigating 'unauthorized access' of its Mythos cybersecurity tool.

What’s Happening

The prediction market assessing whether an Anthropic Claude model score at least 45% on Humanity's Last Exam currently sits at a 14% probability for "YES" and 86% for "NO," reflecting deep skepticism about near-term AI performance on this rigorous benchmark. Humanity's Last Exam is designed to test the outermost limits of AI reasoning, requiring models to solve novel, expert-level problems across mathematics, science, and logic. The low probability comes amid a turbulent period for Anthropic, which on April 22, 2026 confirmed it is investigating a report of "unauthorized access" to its Claude Mythos cybersecurity model through a third-party contractor portal. This security incident has raised questions about whether the company's most advanced systems—potentially including those that might be submitted for the exam—are stable and secure enough for public evaluation. [Engadget, Apr 22]

The controversy surrounding Claude Mythos, which Anthropic has described as too dangerous for public release due to its advanced cybersecurity capabilities, has drawn sharp criticism from competitors. On April 22, 2026, OpenAI CEO Sam Altman accused Anthropic of using "fear-based marketing" to sell Claude Mythos, comparing the strategy to "dropping a bomb while selling a $100 billion bomb shelter." This public dispute highlights the intensifying rivalry in the AI arms race, which accelerated further on April 24, 2026 when OpenAI unveiled GPT-5.5 and China's DeepSeek released a preview of its V4 model—just a week after Anthropic introduced Claude Opus 4.7. The question of whether an Anthropic Claude model score at least 45% on Humanity's Last Exam is now entangled with broader debates about model safety, access, and the pace of capability advancement. [Business Insider, Apr 22]

Looking ahead, the outcome of this market will depend on whether Anthropic can stabilize its development pipeline and produce a model capable of meeting the exam's stringent threshold. The unauthorized access incident, which Bloomberg reported involved a "handful" of users in a private online forum, may delay the release of future Claude iterations or force Anthropic to tighten security protocols, potentially slowing progress. Meanwhile, the rapid release of competing models from OpenAI and DeepSeek raises the stakes for Anthropic to demonstrate that its technology can achieve a 45% score on Humanity's Last Exam—a benchmark that no current AI system has publicly passed. The company has not yet announced which specific Claude model, if any, will be submitted for the exam, leaving the market to weigh the impact of recent security breaches and competitive pressure on its ability to deliver. [Guardian, Apr 22]

Traded on Polymarket — $133K Volume

Polymarket prices this at 14c YES with $133K in volume. Moderate liquidity — use limit orders for positions above $1K to avoid moving the price.