What Actually Changed in the Age of AI This Year, and What Did Not?

Twelve months of AI headlines produced two competing stories. One says everything changed. The other says nothing did. Both are lazy. The honest version requires separating the parts of this year that were genuinely new from the parts that were just louder versions of arguments we were already having in 2025.

The models got better, and the gap between them got smaller

Any launch count, price, benchmark score or “frontier” label written down here would be stale before you read it. Current model catalogs are the source of truth for what exists and what it costs, not a year-in-review list, including this one.¹²³

What that timeline actually tells you: the release cadence compressed, but the capability jumps within each release got smaller. Nobody shipped a model this year that made last year's flagship look like a toy. They shipped models that were reliably, incrementally better at coding, reasoning, and context length, at lower prices. That's not nothing, a frontier model at $2 per million input tokens was unthinkable two years ago; but it's evolution, not the discontinuity the marketing copy implies.

The labour market: bifurcation, not collapse

The confident labour-market story about AI is assembled from vendor, academic, index-company and consultancy analyses covering different geographies with different definitions. Presenting its unemployment, headcount, productivity, wage and age-cohort figures as a single settled labour-market result would mean blending measurements that were never designed to be added together.

This is where the year actually produced something new: enough labor-market data to stop guessing. Anthropic's own economic index, an IMF study of Denmark, and the Stanford AI Index 2026; three independent methodologies; all found no detectable rise in aggregate unemployment among highly AI-exposed workers since ChatGPT's 2022 launch. That result surprised a lot of people who'd assumed mass displacement was already visible in the topline numbers. It isn't.

But aggregate numbers hide a real bifurcation. Of the S&P Global 1200 index, 83% reported lower headcount in January 2026 than January 2025, and globally, job losses tied to AI adoption outpaced job gains by five percentage points over the past twelve months. PwC's 2026 Global AI Jobs Barometer found productivity growth running 40% higher at AI-exposed companies, and wages in AI-"professionalised" roles growing 42% faster than in roles AI merely "democratised." Age matters too: employment in AI-exposed job categories declined for younger US workers over this period while holding steady or rising for older workers already in senior roles.

The technology isn't eliminating work in the aggregate. It's redistributing who captures the upside, and the redistribution is running against anyone starting a career from zero.

That's the real story of the year, and it's more useful than either "AI took all the jobs" or "AI didn't affect employment at all." Both headlines are technically supportable with a cherry-picked chart. Neither is true on its own.

The regulatory clock started running

The EU AI Act applies in stages. The exact obligation date depends on system category, operator, placement date, and provision; use the Commission implementation timeline and regulation rather than one “bulk obligations” sentence.⁴

August 2, 2026 is important for multiple provisions, including specified transparency and Annex III high-risk rules, but exceptions and later dates apply. Penalties also vary by violation, so the single “€15 million or three percent” figure that gets quoted is only one tier of several.⁴

Descriptions of the Colorado law circulate widely in summaries; the primary state text does not support them as written. California AB 2013 requires specified training-data documentation on or before January 1, 2026, while SB 942’s enacted text states an operative date of January 1, 2026; not August 2.⁵⁶

What didn't change: reliability

Cross-model hallucination percentages are not comparable without a shared task, scoring rule, retrieval setting and version; which is exactly what the confident 2024-to-2026 improvement ranges are missing.

So: better models, cheaper tokens, a labor market quietly sorting winners from losers by age and role rather than industry, and a regulatory clock that finally started running instead of just ticking in draft form. What stayed exactly where it was a year ago is the thing everyone building on this technology actually depends on, knowing, reliably, whether the answer in front of you is true. That gap is smaller than it was. It is not closed, and nobody credible is claiming it will be by next year's version of this same article.