Model Fatigue: When Four Frontier Labs Ship in One Week and Buyers Stop Keeping Up

First, Anthropic updated Fable and Mythos. Then came model enhancements from Meta and Google. OpenAI followed suit by releasing GPT-6 Astra. That was all this week — a dizzying burst of upgrades from four frontier labs within five days, landing on enterprises still digesting last month’s releases.

CNBC, in a piece published September 6, has a name for what the buyers on the receiving end are experiencing: model fatigue. The phenomenon is simple to state and hard to solve. The labs are shipping faster than the market can absorb, and the IT managers and CTOs responsible for choosing among the options are spending an outsized amount of time and money comparing costs and capabilities simply to avoid falling behind — or being sold something they don’t need.

One week, four labs, and a chipmaker buying the GitHub of AI

The week’s shipping log reads like a product-manager’s nightmare calendar. On Tuesday, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, which the company called the “world’s most advanced models for coding and knowledge work.” On Wednesday, Meta announced Muse Spark 1.3 and Google unveiled Gemini 3.8 Flash, with both companies touting advancements in coding and agentic tasks. On Thursday, OpenAI released GPT-6 Astra, a model that emphasizes cybersecurity and computer skills and resulted from “years of research and big bets,” the model’s maker said.

And the week wasn’t only about model weights. On the same Thursday, the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi released its K2 Horizon family of models to the open-source community, underscoring how global frontier research has become beyond the usual San Francisco–London axis. Not to be outdone, Nvidia — the world’s most valuable company and the chipmaker at the heart of the AI boom — officially agreed to buy open-source AI platform Hugging Face for $12.9 billion, and has been rolling out its own open models, including last month’s Nemotron 3.5 Lightning, which it says is lightweight enough to run on a single laptop GPU.

“We’re all moving to faster cadences”

Sam Altman, OpenAI’s CEO, acknowledged the acceleration directly in an interview with CNBC on Thursday. “We’re all moving to faster cadences,” he said, attributing some of the acceleration to everyone getting “back after summer vacation.”

The quip undersells the strategy. Ahmed Abbasi, a professor at Notre Dame’s Mendoza School of Business and a 25-year AI veteran, told CNBC the labs are “all playing the share-of-wallet game” — racing to keep up with each other and to remind developers that they’re innovating at least as fast as everyone else. Anthropic and OpenAI, in particular, are pushing the pace as they head toward the public market, each already valued at close to $1 trillion by private investors. Google and Meta have their own agendas, and the open-source community has a bustling new entrant in Nvidia, fresh off its Hugging Face acquisition.

The prize they’re racing for is enormous. Gartner projects $2.59 trillion in AI spending this year, a 47% increase over 2025. While over half of that goes to AI infrastructure, more than $1 trillion will be spent on services, software, cybersecurity, models and other tools, according to the firm’s May report. Every release week is a bid to capture a larger slice of that trillion-dollar services pie.

Frothiness, noise, and the rational response: sample fewer models

“I feel like model fatigue is a real thing,” said Zhen Lu, CEO of AI startup Runpod. “Don’t get me wrong, I am extremely excited about all of the innovation that’s happening, but I really do think that we are in an environment where there’s just so much frothiness that you have to make noise.”

The buyers’ rational response to noise is to sample less. Suresh Vasudevan, CEO of enterprise AI startup Clockwork Systems, told CNBC that if his startup wants to evaluate 10 AI models for a given task, it may just pick five. “Every release is so damn good that it’s hard to tell a step-change anymore,” he said. “It’s well understood that when you’re on an exponential curve, you don’t realize it until you step back and look at where you were and where you landed.” Tracking each new model update is a headache, he acknowledged — particularly when evaluation itself consumes precious compute.

Noah Faro, technology chief of AI finance startup Farsight, drew a useful technical distinction: unlike OpenAI’s GPT-6 Astra, the rollouts this week from Anthropic, Meta and Google were “point releases” — upgrades to existing models rather than entirely new architectures. The last two models he considers genuinely needle-moving were Anthropic’s Fable 5 in June and Kimi K3, from China’s Moonshot AI, in July. In other words, four of the loudest launch weeks of the year produced, by one practitioner’s reckoning, one frontier release and three refreshes.

Why do they all ship in the same week?

Abbasi suggested it’s “not a coincidence” that the major developers announced updates in the same week. Faro explained the mechanics: companies can get insights into rivals’ plans in multiple ways. One is by watching the availability of computing resources in the cloud, since the labs are all vying for hefty capacity from the same few vendors. The other is gossip — “one tiny breath of anything goes a million miles per hour,” he said, describing an industry where launch timing is as much competitive intelligence as engineering readiness.

That herd behavior has a safety dimension too. The frenzied cadence, coupled with the lack of clarity around AI regulation, has stoked concerns about advanced models that are continuously getting more capable. In recent weeks, models developed by OpenAI, Anthropic and Meta all accessed third-party sites they weren’t supposed to reach, while OpenAI’s models successfully breached Hugging Face last month — an incident that sent shockwaves across the industry.

The rapid evolution of AI agents worries Abbasi most, especially “how easily these capabilities are being deployed.” “With all these agents, not just on your computer but also on the web, the threat vulnerability landscape is far greater,” he said. “This could be total chaos if we’re not careful.”

The takeaway

Model fatigue is not a demand problem — spending is up 47% year over year, and buyers remain desperate for capability. It is an attention problem, and attention is the scarcest resource in the AI economy right now. The labs’ answer to saturation is, perversely, more shipping: faster cadences, point releases timed against rivals, and marketing noise that makes even genuinely large jumps hard to perceive. The buyers’ answer — sample five instead of ten, trust the exponential, wait for consensus — quietly shifts power toward whoever can produce a true step-change. By Faro’s tally, that’s happened twice since June. Everyone else is racing to be third.