In the week spanning late August and early September, several labs released new models back to back: Anthropic updated Claude Fable 5.1 and Mythos 5.1, Meta launched Muse Spark 1.3, Google rolled out Gemini 3.8 Flash, OpenAI released GPT-6 Astra, and Nvidia put out Nemotron 3.5 Lightning. In a report on September 6, CNBC gave this week's enterprise-side reaction a name: model fatigue.
The fatigue isn't about capability itself. According to the report, the things executives and IT leaders now have to keep comparing go beyond benchmark scores: price, speed, reliability, and whether a model actually saves labor once it's plugged into their own workflows. Every new batch of models means redoing that comparison from scratch. One company leader interviewed said that of ten new models, he only had time to actually evaluate five.
Spending keeps climbing
Gartner's forecast for global AI spending this year is $2.59 trillion, up 47% from 2025. More than half of that flows to infrastructure; the rest — services, software, cybersecurity, models and related tools combined — adds up to more than $1 trillion.
Break that $1 trillion down, and model API fees are just one line item. What model selection actually eats up is evaluation environments, internal benchmarks, migration work and regression testing — the things that never show up on a price quote. A mid-sized company that re-evaluates its primary model every quarter runs through four rounds a year; each round involves rewriting prompts, rerunning eval sets and coordinating with everything downstream, and by headcount that work often costs more than the API bill itself. This is a rough estimate based on public pricing and common engineering practice — no company has disclosed its own model-selection costs.
Why the supply side can't slow down
From the model makers' side, the rapid-fire releases are a fight for share of wallet. That's the term Ahmed Abbasi of Notre Dame's Mendoza College of Business uses: labs not only have to keep pace with each other, they have to keep proving to developers that they haven't fallen behind. Both Anthropic and OpenAI are moving toward public markets, and private-market valuations for both are approaching $1 trillion — the release cadence and the fundraising narrative are running on the same track.
The other end of that track is buyer attention. The faster models turn over, the more likely enterprises are to delay their selection process, or simply lock in one vendor and stick with it — either reaction blunts the impact of a new release. CNBC's report doesn't cite a quantitative survey; the claim comes from how the people it interviewed described feeling.
What can eventually be checked against reality comes later. The $2.59 trillion figure is an annual forecast from a research firm, and the actual numbers won't be in until next year; whether enterprises end up switching models faster or holding off will only become clear once cloud providers' quarterly earnings break out AI service revenue. For now, the only thing that's certain is the pace of releases.
Sources: CNBC, CocoLoop, Gartner forecast data, Digital Today; the $2.59 trillion spending forecast and 47% growth figure, the infrastructure share and $1 trillion services-and-software figure, and the list of five model releases that week were each checked individually.