Sup AI tops Humanity's Last Exam

This is a strong benchmark signal, but it is also a product-positioning move: Sup AI is selling orchestration quality, not a single magic model. The result is interesting because it leans on ensemble diversity and confidence filtering, which is a more defensible story than “our model is smarter.”

• –The headline number matters: 52.15% on HLE is positioned as 7.41 points ahead of the next best model in its setup.
• –The benchmark run used web search and custom prompts, so it is not a clean apples-to-apples comparison with raw model scores.
• –The product itself looks closer to an accuracy-first research assistant than a general chatbot, with source transparency, file search, and context compaction as core features.
• –If the claim holds up outside the benchmark, the real moat is routing and verification logic, not model ownership.
• –The risk is obvious: ensemble systems can look great on curated evals while still being hard to trust on messy real-world workflows.