BLOG
Big Models Update Four Times a Week, Enterprises Are Suffering from 'Model Fatigue': Chase, or Lie Flat?
Opening: The Media Coined a New Term — “Model Fatigue”
Over the past week, the large model market has seen a dense wave of new releases: on September 1, Anthropic rolled out two new models in one go; on September 2, Meta and Google launched theirs on the same day; on September 3, OpenAI unveiled GPT-6 Astra — four major players making consecutive announcements in a single week.
When Guancha.cn reported on this, it used a new term: model fatigue. It’s not that users are losing interest in AI, but that new models are being released so quickly that developers, enterprise customers, and investors are all struggling to keep up — you haven’t finished evaluating one model before the next version is already out; you’ve just completed deployment, and you might already be facing an upgrade.
One number really drives the point home: the median release interval for major models has shrunk from 37.5 days in 2023 to just 11 days today.
I went through this report and the industry logic behind it, then posed the question to Yongliang — a 17-year software industry veteran with 7 years of AI experience, currently serving as an AI technical director. He has personally built enterprise-grade AI efficiency solutions, and “model selection” is exactly his day-to-day work.
Shiwen: Four updates in a week, even professionals like you can’t keep up with the pace. How does it feel? Yongliang: To be honest, I’ve stopped chasing. It’s not lying flat — I’ve figured something out: chasing models is an infinite game. Enterprises can’t win it, and they don’t need to. Shiwen: Alright, then let’s talk it through: why is it happening so fast, what exactly are enterprises worried about, and what should they do?
Q1: How Dense Is This “Weekly Update” Cadence? What Exactly Is “Model Fatigue”?
Yongliang: Lay out the timeline, and you’ll see how crazy it is.
On September 1, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, with the latter billing itself as “the world’s most advanced coding and knowledge work model.” On September 2, Meta released Muse Spark 1.3, and Google released Gemini 3.8 Flash. On September 3, OpenAI released GPT-6 Astra, calling it “the world’s smartest and best-aligned model.” On the same day, MBZUAI in the UAE also open-sourced the K2 Horizon series.
One week, five waves. And this isn’t a one-off — Google updated Gemini Flash four times in 106 days, with the latest generation coming just three weeks after the previous one. OpenAI’s own median release interval has compressed from 170 days in 2023 to 49 days this year.
Where does the “fatigue” come from? The original quotes from several practitioners interviewed by Guancha.cn are very telling. Runpod’s CEO said “model fatigue is real,” and Clockwork Systems’ CEO put it more bluntly: every release is “unbelievably good,” to the point where it’s hard to tell what counts as real leapfrog progress — they all sound great, so which one do you actually pick?
For enterprises, this translates to real costs: every time a new model comes out, you have to re-evaluate performance, price, compute requirements, security, API compatibility, and fit with existing business. This process used to happen once or twice a year; now it happens four or five times a quarter. The most striking detail in the report: one enterprise had 10 candidate models to evaluate, but could only do 5 — they didn’t have the resources for all of them. Small and medium-sized enterprises have it even worse, probably can’t even finish 5.
Q2: Why Are Vendors Releasing Models Like Crazy?
Yongliang: Three driving forces are stacking up, and none of them allow for slowing down.
First, the baton of the capital market. Both Anthropic and OpenAI are preparing for IPOs, with private valuations approaching $1 trillion. What does this valuation narrative need? It needs a “continuous growth story.” Releasing new models constantly is the most direct signal that “we’re still leading.” Slow down for one quarter, and the valuation story takes a hit.
Second, the prisoner’s dilemma of “capability convergence.” Gartner coined a concept called capability convergence back in June this year — when more and more models perform similarly on benchmarks, the advantage you build by being “first to release” gets erased by others in a short time. This leads to a vicious cycle: the shorter the advantage retention period, the more you have to rely on release frequency to maintain presence. No one wants to release this fast, but not releasing is equivalent to dropping out of the race.
Third, the catch-up by Chinese vendors has changed the game. The report mentions that open-weight models from Chinese companies like Moonshot AI and Zhipu AI have performance approaching or even partially reaching closed-source levels, while holding a clear edge in cost and openness. Data from a 21st Century Business Herald report is even starker: the share of tokens used by US companies calling Chinese models rose from 4.5% in the first half of 2025 to over 30% after February this year; for the same workload, Claude charges $25 while DeepSeek only charges $0.18. When the price-performance gap is a hundredfold, Silicon Valley giants can only offset “more expensive” with “faster updates.”
So you see, this is a race no one dares to hit the brakes on. OpenAI’s CEO Sam Altman himself has admitted that the entire industry is “accelerating the pace.”
Q3: What Exactly Are Enterprises Worried About? How Do You Calculate the Cost?
Yongliang: The worry isn’t that models aren’t good — it’s that “good” has become too cheap.
I’ve built AI solutions for enterprises, and I’ve run this calculation many times. On the surface, models getting stronger and cheaper is a good thing. But for people doing model selection, it means three things:
First, evaluation costs have become a recurring expense. Selection used to be a one-time project; now it’s a constant job. You have to maintain evaluation sets, run tests, compare results — this whole process has to be repeated every time a new model comes out. At many companies, the AI team spends half their time “chasing updates” and the other half “regretting their choices.”
Second, the risk of lock-in has grown. If you tie your business deeply to the features of one vendor’s model, and they release a new version next month with changed behavior, your application might break. The faster models update, the higher the risk of “locking into one vendor” — but the maintenance cost of “supporting every vendor” is also right there.
Third, FOMO has turned into decision paralysis. There’s always someone on the team asking “XX released a new version, should we switch?” Switching means re-testing and re-adapting; not switching means you can’t stop second-guessing. This constant back-and-forth erodes trust and morale more than technical costs do.
Gartner’s judgment in the report is actually a wake-up call for all enterprises: the advantage of foundation models is becoming temporary. Translated into plain language: the era where you could “pick the right model once and be set forever” is over.
Q4: Chase or Lie Flat? Tell Us About Your “Respond to Change with Constancy” Approach
Yongliang: Neither is right. The correct answer is: turn “chasing models” into an architecture problem, not a manpower problem.
When I build enterprise AI solutions, I set one iron rule: no business directly depends on the version number of any vendor’s model. It breaks down into three layers:
First layer: model gateway. All calls go through a unified gateway, which connects to multiple models underneath. Business code only recognizes the interface, not the model. Switching models is a one-line config change, not a month of code changes. This way, the impact of “weekly model updates” on you drops from “project-level” to “configuration-level.”
Second layer: automated evaluation. Don’t rely on human intuition to judge “is the new model good.” Build your own evaluation set — use your real business inputs and outputs as the benchmark. When a new model comes in, run an automated evaluation, get scores, get comparisons. Build this thing once, and it benefits you every time. Your evaluation set is more accurate than any third-party leaderboard, because it’s fed with your own business.
Third layer: value sinking. This is the most important one. Gartner says value is shifting downstream — to data quality, integration depth, and domain fine-tuning. This aligns completely with my experience building solutions: what really builds a moat is never which model you use, but the data you’ve accumulated, the processes you’ve connected, the domain adaptation you’ve tuned. Models are rented capability; data and processes are your own assets.
One sentence summary: Chase, and you’ll never finish chasing; lie flat, and you’ll be out of the game soon. The only sustainable approach is to make “switching models” cheap.
Q5: Gartner Says Value Is Shifting Downstream — Where Should Individuals and Enterprises Build Up?
Yongliang: Three layers of accumulation, ranked by how “unremovable by model changes” they are.
First, data. Your business data, user feedback, domain knowledge base — these are things models can never take away. The most important thing enterprises should do right now is to govern, structure, and make searchable the data scattered across departments. No matter how strong a model is, if it can’t access your data, it’s useless to you.
Second, processes and integration. How AI fits into your business processes, who reviews outputs, how to handle failures — this “integration depth” is the second moat. That’s why I always say enterprise AI efficiency isn’t about buying tools, it’s about getting the processes right.
Third, judgment. For individuals, the shelf life of “knowing how to use a certain model” is only a few weeks, but the shelf life of “being able to judge whether AI output is good” is an entire career. Models update every 11 days — you’ll never keep up chasing them. But the ability to judge what should use AI, what shouldn’t be trusted, and what to do when things go wrong — this ability becomes more valuable the stronger models get.
Back to the original question: what to do about model fatigue? My answer is already out there — don’t tie your company to a model’s version number; sink your value into things models can’t take with them. Vendors are playing an infinite game. You just need to guard your own moat and let them do the racing for you.
Closing
Shiwen: Finally, sum up this episode in one sentence? Yongliang: Vendors race on models, you race on data and processes — models are rented, your moat is your own. Shiwen: This sentence is for everyone. See you next time.
[Technical Deep Dive] What Exactly Does “Model Gateway + Evaluation Set” Look Like?
We kept saying “turn chasing models into an architecture problem” in this episode, so the technical form of this setup is worth expanding on.
Model Gateway (AI Gateway) is essentially a reverse proxy: business parties uniformly call the gateway’s API, and the gateway is responsible for routing to OpenAI/Anthropic/DeepSeek/local models, while also handling key management, usage statistics, rate limiting, and failover. Open-source solutions like LiteLLM and OneAPI are already very mature, and cloud vendors also have managed versions. Its value has been dramatically amplified in the “weekly update era”: changes at the model layer are absorbed by the gateway, and the business layer feels nothing. If a vendor’s model goes down or acts up, switching traffic is just a config change.
Evaluation Set (Eval Set) is a set of “standard test questions” for your own business: real inputs + judgment criteria for expected outputs. When a new model comes in, you run the test questions through it and automatically score and compare results. The key is the words “real business” — third-party leaderboards measure general capability, while your evaluation set measures “does it work in your specific scenario.” A medical Q&A system and a marketing copywriting system have completely different definitions of “good,” and general leaderboards can’t answer that for you. The evaluation set doesn’t need to be huge — dozens to hundreds of high-quality samples are enough. The difficulty isn’t in quantity, but in writing “what counts as a good output” into rules that machines can judge.
These two things together form the stabilizer for enterprises in the “weekly model update era”: the gateway handles “switching fast,” and the evaluation set handles “switching right.” Add to that externalizing prompts, business rules, and domain knowledge into versioned assets (instead of scattering them in code), and your AI system transforms from “an application tied to a specific model version” into “a chassis that any model can plug into.”
This is also the ultimate antidote to “model fatigue”: it’s not about catching up with all models, but about making the cost of “not chasing” bearable.