#benchmarks

A new benchmark called CiteVQA shows that leading AI models frequently give accurate document answers while failing to identify the actual supporting passage, a gap researchers call attribution hallucination.

A new benchmark called SOOHAK pushes frontier AI models beyond standard competition math by testing research-level reasoning and whether systems can recognize when a problem has no solution at all.

Researchers at Carnegie Mellon University say a new benchmark measuring real V8 exploitation found Anthropic's Claude Mythos Preview well ahead of OpenAI's GPT-5.5, though at far higher cost.

A new academic benchmark shifts attention away from cinematic polish and toward whether AI video models can continue scenes in ways that obey physics, social norms, and basic logic.

A large study of more than 34,000 real-world agent skills suggests the modular instructions praised in benchmark settings deliver far smaller gains when models must find and apply them on their own.

A new benchmark built to test whether multimodal AI systems know when to ask for more information shows that most current models still guess, hallucinate, or refuse instead of requesting the visual context they need.

A new analysis found Google’s AI Overviews answered benchmark questions correctly about nine times out of 10, but the remaining error rate could still translate into a vast volume of wrong answers.

Apple's new entry-level MacBook Neo with 512GB unified memory delivered benchmark results comparable to powerful cloud instances in DuckDB database workloads, raising fresh questions about the economics of cloud computing.

Early Geekbench results for the iPhone 17e's A19 processor reveal notable single-core and multi-core performance improvements over its predecessor.

Anthropic's latest mid-tier model, Sonnet 4.6, debuts with record scores in software engineering and computer use benchmarks, plus a doubled context window of one million tokens. The release becomes the new default for free and pro users.