Where Is AI Today?

AI Is Powerful, but Uneven

In 2026, leading AI systems can write, code, analyze documents, work with images and audio, use tools, and solve some extremely difficult academic problems. At the same time, they can still make surprisingly simple mistakes.

Researchers sometimes call this a jagged frontier: capability is very high in some directions and much lower in others.

Language, Coding, and Reasoning

AI is especially strong with language and increasingly strong at software work. Stanford’s 2026 AI Index reports that performance on SWE-bench Verified rose from about 60% to near 100% in a year, while frontier systems also meet or exceed human baselines on several difficult science, multimodal, and mathematics benchmarks.

That does not mean every real software project or reasoning task is solved. Reliability remains a separate problem.

Agents Are Moving From Answers to Actions

AI agents can use computers and tools to complete parts of larger tasks. On OSWorld, a benchmark of real computer tasks, leading performance rose from roughly 12% to 66.3%—a huge improvement, but still a failure on about one out of every three tasks.

The Physical World Is Harder

Robotics shows the unevenness clearly. Stanford reports roughly 89.4% success on one simulated manipulation benchmark but only about 12% success on tested real household tasks.

A home contains slippery objects, strange lighting, clutter, people, pets, doors, drawers, and surprises. The physical world is difficult.

Capability vs. Reliability

A system may be able to do a task without being reliable enough to trust it every time. That distinction matters most when mistakes affect health, money, safety, or other people.

Where We Are