What to Expect from AI in 2026

Last year was a jam packed with model iterations, massive investments, mergers, and Tech CEO’s making wild claims.

In 2026, expect fewer toys but more promises. The industry is running into the problem of physics, economics, and reality.

First, the cadence of meaningful frontier model releases will slow. The labs will keep pushing, but it will be more bark and less bite. With today’s evaluation regimes, the easy wins have largely been banked. New checkpoints will still arrive, and the press releases will still speak in superlatives, yet many updates will amount to small deltas: marginal gains on saturated benchmarks, modest cost reductions, slightly better latency. These improvements are still welcome, but it won’t justify weekly talk of imminent “general” intelligence.

Second, a reckoning is coming for “AI agents”. The pitch was seductive: autonomous systems that browse, plan, act, and learn continuously, replacing whole swathes of the workforce. But today’s agentic systems demand a great deal of bespoke data, careful engineering, and constant monitoring (something very few people have). They also fail in ways that make you facepalm, where a primary school kid would not. Many agent-first startups will try to pivot but likely evaporate, having discovered that demos are cheap and operations are not.The more durable winners will look like well-designed tools. Directed, practical human-in-the-loop applications will reach product market fit in growing numbers. Instead of promising a digital employee, they will offer measurable assistance in tightly scoped tasks: drafting, triage, search, summarisation, classification, anomaly detection, and decision support where accountability remains human. A relatively small cohort of applications will emerge as genuine value adds, because they are built around reality rather than novelty.

Meanwhile, the big labs will begin to change track. After predominantly engineering-led scaling, attention will shift back to research questions that remain to eluded humanity, what is general intelligence? How can today’s AI pass the Turing test yet can’t perform basic tasks? Expect confident headlines that suggest continuous learning is just around the corner but do not expect anything tangible in this space for 2026.

Finally, open source will use this time to reduce the gap. This will be great for competition, privacy, and sovereignty. The state-of-the-art models will likely still belong to the large labs due to their access to compute, but there may be some “self-hostable” competitors by the end of the year.

If you take one lesson into 2026, let it be this: put a personal deadline on that fully autonomous agent (they’re the self-driving car of the workforce and already a broken promise). Instead, pay attention what is working in practice. Constrained systems with clear scopes, human oversight with incentives aligned to real outcomes.

‍

12 January 2026

‍