An honest look at where AI actually stands in April 2026, and why the MIT Project NANDA study reported that 95% of enterprise AI pilots produce no measurable return.

Frontier capability has moved fast: Claude Opus 4.7, GPT-5.4 with native computer use, a Chinese open-weight tier within shouting distance of the US labs, and METR time-horizon numbers that put top models at 3-5 hours of effective autonomous work. But the binding constraint on AI for businesses right now is not capability. It is reliability. And reliability is moving far slower than the capability curve suggests.

Part 1 of a two-part series. In this episode:

The real state of the frontier in April 2026 (Opus 4.7, GPT-5.4, Gemini 3.1 Pro, Chinese open weights)
Why METR's "time horizon" metric matters more than one-shot benchmarks
The three forces driving current capability gains (compute scaling, post-training, interpretability)
The limits that do not show up in keynote demos: hallucination rates, long-horizon reliability, the gap between benchmark and real work
The MIT Project NANDA study everyone quotes, and what it actually found
The circular-capital dynamic: $700 billion in 2026 hyperscaler capex and the money flowing between AI suppliers
Why the next binding constraint on AI infrastructure is electrical power, not chips

Full written analysis: https://prentusai.com/news/the-state-of-ai-april-2026.html

Part 2 covers geopolitics, the labor picture, the next 12-24 months, and a practical roadmap for a business owner deciding what to do next.

Agents at Work is the podcast of PrentusAI, an independent learning hub covering OpenClaw, NemoClaw, and AI agent technology in plain English for business owners. Subscribe for weekly deep dives.

Web: https://prentusai.com Newsletter: https://prentusai.com Full article archive: https://prentusai.com/news/

State of AI, Part 1: Why 95% of AI Pilots Fail