Choosing an AI model is no longer about picking “the best”
At ePrize, we are currently reassessing the AI models we use in development, shifting our primary stack from Claude and ChatGPT to a combination of GLM 5.2 and Kimi K3.
One point worth clarifying upfront: these are two models with distinct personalities. GLM 5.2 is a model developed by Zhipu AI (Z.ai) in China, and Kimi K3 is a model from Moonshot AI, also based in China. Kimi K3 launched as a 2.8-trillion-parameter open-weight model, explicitly positioned as offering “capability that is sufficient — or better — for real work, at a dramatically lower cost” than existing high-performance, high-cost frontier models such as Claude Opus 4.8. On pricing, Kimi K3 is listed at $3 per million input tokens ($0.30 on cache hit) and $15 per million output tokens, versus Claude Opus 4.8 at $5 / $25. Even on output price alone, Kimi K3 sits at roughly 60% of the frontier model.
On the performance side, “a step below the top model” is too simple a framing — the picture shifts depending on the benchmark. In evaluations by the independent organization Artificial Analysis, Kimi K3 ranked 4th among major frontier models on the composite score, reportedly ahead of Claude Opus 4.8. At the same time, on some indicators such as practical code-fixing ability, it falls slightly short of the top models in certain scenarios. The area drawing the most attention is its strength in front-end development: it took the top spot in the Frontend Code Arena, an evaluation based on blind testing. In other words, on certain task types it matches or exceeds the top models, and especially for agentic use cases and coding work, Kimi K3’s strengths stand out — that is the picture closer to reality.
What matters is that this kind of model — “slightly held-back on performance but massively lower in cost” — is not unique to Kimi K3. OpenAI, Anthropic, and other vendors are also releasing a stream of models with the same intent. This is therefore not a single company’s strategy but a trend happening across the industry, and should be understood as such.
GLM 5.2 is also a model positioned within this same trend, like Kimi K3. As we discuss in the next section, Chinese-origin models are rapidly expanding their global usage share, armed with open-weight releases and overwhelmingly low prices. GLM 5.2 and Kimi K3 can both be read as “a low-cost, high-performance alternative to US frontier models” — a movement sharing that common angle.
Hearing this, some readers may feel “won’t performance drop?” or “is a Chinese model safe?”. But when we actually used them, we often found quality on par with — or, depending on the task, above — the top models. And more importantly, this is not just our movement. Right now, even in US-based companies, a shift toward lower-cost models and Chinese-origin models is spreading rapidly, driven by rising costs.
In this article, we want to organize why this reassessment is happening, and what the truly important perspective is when selecting AI models — together with the cost issues we have repeatedly covered in past ePrize articles.
A second current against “high performance, high cost”
The first thing to hold on to is that models like Kimi K3 — “sometimes above the top model depending on the task, with enough overall performance delivered at low cost” — are becoming an industry-wide trend.
Behind this is the fact that AI use has shifted from “experimental stage” to “the stage of being used heavily in real work”. When agents run autonomously and execute tasks repeatedly, the cost per task stacks up and directly drives monthly AI spend. The value of models that are “not the absolute best on single-shot accuracy, but can run a huge volume of tasks at low cost” is therefore rising fast.
In response, major vendors such as Anthropic, OpenAI, and xAI are also releasing models that pitch high cost performance. A pattern is spreading across the working floor: reserve the top models for high-stakes judgments, and deploy cost-efficient models for everyday, routine tasks.
Chinese-origin AI models rising fast — US companies are switching too
From here, let us look at the other current: the movement of Chinese-origin models.
Let us begin with objective data. According to aggregates from an API integration platform for AI models, usage of Chinese-made AI models has expanded sharply in 2026, eventually surpassing usage of US-made models. Models from Chinese companies now occupy many of the top slots by call volume, and the reasons cited behind that are overwhelmingly good cost performance, high inference efficiency, and strong long-context handling.
US companies are visibly switching as well. The CEO of a major cryptocurrency exchange has publicly stated that AI costs were halved. There are also reports that on a developer platform, the share of usage going to Chinese-made AI is approaching 50%. On pricing, there are cases where the price is literally a few-dozen-times lower than that of top US models.
Of course, part of this is driven by temporary supply instability on the US side due to export controls and similar measures. But bigger than that is a simple economic rationale: “equivalent or near-equivalent performance, at overwhelmingly lower cost”.
“A Chinese model” and “AI operated in China” are different things
Here, we want to untangle one commonly misunderstood point. People sometimes say “Chinese-made means dangerous” or “avoid Chinese-made”. That is a somewhat coarse argument.
What matters is not the origin of the model itself, but where and how that model is operated. For instance, it is now common to take a model that a Chinese company has published as open weights — freely modifiable and commercially usable — and host it on a US-based cloud platform for use. In that case, even if the model’s weights originate in China, the actual data processing and operations take place on infrastructure you manage, or on a platform you trust.
In other words:
- The origin of the model
- The where and who of operating that model
These two must be considered separately. For work involving highly confidential data, the design of the operations side is critical — but that is essentially a different question from “avoid everything Chinese-made”. In practice, teams that can choose their own models are settling into an approach of running the same task in parallel across multiple models, measuring unit price and output quality empirically, and then deciding on adoption. Not judging by nationality alone, but judging by empirical measurement. This will become the baseline stance for model selection going forward.
The meaning of “top of the benchmark” is fading
Another thing we feel strongly is that the meaning of performance comparisons — “which model is 1st or 2nd” — is itself thinning out.
Every time a new model is released, benchmark scores make headlines. Indeed, in 2026 we have seen models arrive claiming to deliver Claude-Opus-class performance faster and at lower cost, and competition among vendors is intensifying. Some models beat the existing top model on specific benchmarks while falling short on others, and we also see models clarifying their stance as “competing on cost performance rather than pure top performance”.
This means that performance comparisons at release time are, at best, “information in that one moment”. A few months later a different model rises and the ranking shifts easily. This kind of information is useful in itself — but I have doubts about making it the basis for vendor selection itself.
I went through a phase where I judged Claude to be the best. But that was always temporary, and what truly matters, I have come to feel, is having the organization to keep choosing the optimal model at any given point.
The invisible risk of vendor lock-in
Every time a new model appears, you compare options and decide which vendor to contract with. This decision process itself actually carries a large hidden risk: vendor lock-in.
If you build a design that depends deeply on a particular vendor’s API or tools, you cannot easily migrate later even when you want to switch to a lower-cost or better-performing model. Prompts, surrounding tooling, internal operating flows — everything gets optimized to that one vendor.
AI evolves very fast, and the landscape can change substantially in months. If you build a structure you cannot easily exit once contracted, you will not only miss out on the benefits of change — you may end up continuing to pay inflated costs.
That is why, going forward, a design that does not depend on any specific model or vendor — in other words, building in the flexibility to switch between multiple models depending on the situation — will become a very important management decision.
The “obsolescence risk” carried by multi-year projects
When planning a system development project in an enterprise, it is not unusual to proceed over a year or more. But given AI’s speed of evolution, that pace makes the risk of the project’s very premises becoming obsolete extremely high.
A system you spent enormous cost and time building may, by the time it is complete, sit in a world where “a cheaper, higher-performance option” simply exists as a matter of course. From here on, this will be a risk factor that cannot be ignored in any project that uses AI.
This is not simply a matter of chasing trends or not. Whether or not you build in, at the design stage of a project, a structure that lets you flexibly switch models and vendors later, will significantly affect the project’s success or failure and its return on investment.
The “use” of AI models is shifting from a luxury good to a commodity
Finally, we would like to broaden the perspective a little.
When you hear the phrase “AI infrastructure”, many readers may think of massive investments in GPUs and data centers. Indeed, investment in compute is often spoken of as a strategic asset that shapes corporate competitive advantage, and “AI infrastructure” in this sense is positioned as a differentiator — a “moat” that is not easy to imitate.
But what we want to discuss here is a different point. What we should be reassessing is that the object of “use” — the AI model itself — is changing in nature from a luxury good to a commodity.
AI models have, to some degree, been treated almost like luxury goods, where “higher performance means higher value”. But as we enter the phase of using AI heavily in actual operations, that positioning is shifting significantly. The model itself will, going forward, become something closer to a daily-use commodity or a public good — used routinely and as a matter of course — much like electricity, gas, and water.
Different logic applies to commodities and public goods than to luxury goods.
- Food, with a few luxury items excepted, is expected to be supplied more cheaply and more widely.
- A car, when used simply as a means of getting around, does not need to be a luxury car.
- Electricity and gas have no concept of “high-performance electricity” or “high-performance gas” to begin with — a fair price forms through the balance of supply and demand.
AI model usage is expected to follow the same path. Instead of “expensive is expected”, the dynamic will be “if it isn’t reasonably priced, people won’t use it widely”. In fact, major vendors such as Anthropic, OpenAI, and xAI are releasing high-cost-performance models in succession in response to this current. It is precisely because we are in a period where each company is fiercely competing to capture market share that these moves are accelerating — that is fair to say.
In other words, there are two layers to the “infrastructural-ization” of AI. One is infrastructure as a differentiator — compute as a strategic asset. The other is the commoditizing commodity side — the use of models. We want to emphasize again that this article deals with the latter: the price competition in model selection and procurement.
Conclusion: What matters is building an organization that can keep choosing
What we should truly keep in mind in AI model selection can be summarized as follows.
- Judge by operational reality, not by model origin — not “it’s Chinese-made therefore”, but where and how it is operated
- Understand that top-of-benchmark is only a momentary piece of information — do not make performance comparisons at release time the absolute criterion for vendor selection
- Avoid vendor lock-in and secure the flexibility to switch — avoid designs over-optimized for one specific model
- Build obsolescence risk into project design — for year-plus plans, leave room to reassess the model partway through
- Treat AI model usage as a commodity and be conscious of cost structure — while compute investment is a differentiator, model usage itself should be premised on a future where “high performance = high cost” is no longer the default
AI technology will keep evolving at furious speed. Rather than being tossed around by that change, the priority is to build a flexible organization premised on change. This, more than anything, will likely become the most important competitive strength for companies using AI in the era ahead.
At ePrize, we will continue to publish practical insights on AI use in this blog, holding these perspectives.