Zoom reports a new SOTA on the benchmark for AI reasoning

By EngineAI Team | Published on December 17, 2025
Zoom reports a new SOTA on the benchmark for AI reasoning
Zoom recently revealed that their "federated" AI system outperformed Google Gemini on one of the most difficult expert-level reasoning exams in the industry, scoring 48.1% on Humanity's Last Exam.

The specifics:

Using a "Z-scorer" selection system, Zoom's federated approach combines its own small models with the best models from OpenAI, Anthropic, and Google.

Zoom's system surpasses the previous leader Gemini 3 Pro's 45.8%, which has since been eclipsed by the newly launched GPT 5.2's 50%.

Zoom promises more precise summaries, reasoning, and task automation throughout the platform, and the system will power the company's AI Companion 3.0.

The AI program DeepWriter, which asserts that it received a score of 50.91% on the benchmark in late November, also contested the assertion.

Zoom—the cutting edge of AI research? Jokes aside, this is an excellent outcome, albeit further evidence is needed. However, a federated strategy that combines several frontier models could be an appealing model for businesses looking to develop new cutting-edge AI capabilities without starting from scratch.