Stanford's Human-Centered Artificial Intelligence Institute published its 2026 AI Index on 13 April, delivering the most comprehensive annual snapshot of the field's trajectory. The headline numbers are striking: global AI investment hit $581 billion in 2025 — more than double the $253 billion in 2024 — with the United States capturing $344 billion of that total. The US released 50 'notable' AI models in 2025, maintaining its lead over China at 30, while industry now accounts for over 90 per cent of notable model releases, up from roughly half a decade ago.
The compute picture is equally dramatic. World AI compute capacity has grown at 3.3x annually since 2022, increasing 30-fold since 2021, with Nvidia commanding over 60 per cent of global capacity. That compute is translating directly into capability gains: on Humanity's Last Exam, top performers including Claude Opus 4.6 and Gemini 3.1 Pro achieved approximately 50 per cent accuracy by April 2026, up from 8.8 per cent a year earlier. Yet the Index also documents persistent blind spots — GPT-5.4 managed only 50 per cent accuracy reading analogue clocks, and Claude Opus 4.6 just 8.9 per cent on the same task.
The environmental cost of frontier training is now quantifiable and sobering. The report estimates that xAI's Grok 4 produced roughly 72,000 tonnes of CO2-equivalent during training — vastly higher than GPT-4's 5,184 tonnes or Llama 3.1's 8,930 tonnes. On the employment front, entry-level software developer and customer support positions declined while mid-career roles held steady, though the authors caution that broader economic factors make it difficult to isolate AI's specific impact. Public sentiment remains ambivalent: 59 per cent of respondents believe AI's benefits outweigh its drawbacks, but 52 per cent report feeling nervous about AI-integrated products.
For context engineers, two findings stand out. First, the benchmark acceleration from 8.8 per cent to 50 per cent on Humanity's Last Exam in a single year demonstrates that the capability frontier is moving faster than most planning horizons assume — the tools we build workflows around today will be materially more capable by the time COR Summit convenes. Second, the trust gap is real: only 31 per cent of US respondents trust their government to regulate AI effectively, the lowest figure globally, despite the US leading in investment. That disconnect between capability growth and governance confidence is the backdrop against which every enterprise AI deployment now operates.
