AI in the first half of 2026, by the numbers

Half a year of AI news produces a lot of adjectives and few comparable numbers. This review keeps only figures that can be traced to a published source: how much is being spent, who is using the technology, how much better the agents have become, what the leading models cost, and what went wrong. Where a figure comes from a vendor about its own product, we say so.

1. Spending: the infrastructure bill keeps rising

In late April, Meta and Alphabet raised their plans and Microsoft gave a new figure. Meta lifted its 2026 capital expenditure range to $125-145 billion, Alphabet to $180-190 billion, and Microsoft guided to about $190 billion for calendar 2026 (including $25 billion tied to higher component pricing). Amazon said its plan, which in January it expected to approach $200 billion, was largely unchanged ^[2]. One analysis put the combined total at about $725 billion ^[2]; the article does not show its arithmetic, but the high ends of the four figures above sum to $725 billion (derived: 145 + 190 + 190 + 200).

The supplier side shows the same pattern. Nvidia reported $81.6 billion of revenue for the quarter ended April 26, up 85% year over year, of which Data Center was $75.2 billion (up 92%) and networking alone $14.8 billion (up 199%). It guided next quarter to $91.0 billion ^[3].

2. Adoption: broad use, shallow change

The Stanford AI Index reports 88% organizational AI adoption and 53% adoption of generative AI across the population within three years; in the United States the figure is 28.3% ^[1]. Gallup's workforce survey (February 2026, 23,717 U.S. employees) found that half of employed adults use AI in their role at least a few times a year, up from 46% the prior quarter. Daily use is 13%, and 41% say their organization has integrated AI tools ^[4].

The depth of change is lower. Only about one in ten employees at adopting organizations strongly agree AI has transformed how work gets done, although 65% say it improved their productivity or efficiency ^[4]. Gallup's separate analysis found that frequent use is much higher where managers actively support it: 78% versus 44% ^[7].

3. Capability: agents got markedly better

Per the AI Index, success on the OSWorld computer-use benchmark rose from 12% to about 66%, and performance on SWE-bench Verified (a coding benchmark) rose from 60% to near 100% in a single year ^[1]. Near-ceiling scores are a signal to be careful: a benchmark that most leading models nearly top no longer separates them. The same report notes that the top model reads analog clocks correctly only 50.1% of the time ^[1], a reminder that strength on one task does not transfer to others.

Model releases tracked this progress. Anthropic released Claude Opus 4.8 on May 28 ^[6] and, on June 9, Claude Fable 5, a model in the "Mythos-class" tier that Anthropic says sits above its Opus class, with Mythos 5, the same model with some safeguards removed, restricted to Project Glasswing cyber-defense partners ^[5]. The company's benchmark claims for Fable 5 are its own and customer-reported; we have not independently verified them.

4. Pricing: frontier price tiers, with room below

Published list prices per million input / output tokens: Claude Opus 4.8 at $5 / $25, unchanged from Opus 4.7 ^[6]; Claude Fable 5 at $10 / $50, which Anthropic says is less than half the price of Claude Mythos Preview ^[5]. Subscription access to Fable 5 was included through June 22, then required usage credits from June 23 ^[5], an example of how access terms, not only list prices, shift quickly.

5. Risk: incidents and access shocks

Documented AI incidents rose to 362 from 233 in the AI Index ^[1]. Availability risk also became concrete: access to two flagship models was temporarily suspended in June ^[5]. Access status for these models may still be changing, so any workload that depends on one model should have a tested fallback.

What the numbers say together

Spending and adoption are both high, while depth of change, measured by employees who say work has been transformed, is low. That gap, not model capability, is where most organizations should focus in the second half: manager support, clear policy and systems that integrate with daily work all correlate with more frequent use in Gallup's data ^[7]. Plan for model access and pricing to change under you, and keep the ability to switch providers.