As AI capabilities become more accessible and cost-effective, companies are prioritising end-user perception over raw model performance, signalling a disruptive shift in AI deployment strategies.
Artificial intelligence is entering a more pragmatic phase. Building capable products is getting easier, but choosing the right model is becoming a more complex commercial judgement. For many teams, the real question is no longer which system performs best on paper, but whether users can actually see enough difference to justify higher costs.
That shift was visible in Razer’s development of Razer AVA Mini, an AI experience built to generate personalised companion images. Quyen Quach, vice president of software at Razer, told KoreaTechDesk that the company compared open and closed image models side by side before making a deployment decision. The team found that open-source Flux-family models met the required quality bar and that users could not perceive a meaningful difference.
The lesson is broader than a single product launch. Benchmark charts remain useful for comparing reasoning, coding, image generation and multimodal capability, but they do not measure perceived value. Once Razer concluded that the visual quality gap was not noticeable to end users, the focus moved to infrastructure efficiency and cost discipline. In other words, model selection became a product and operations decision, not just an engineering one.
That approach fits a wider trend in the market. McKinsey’s latest research on open-source AI found that more than half of surveyed organisations already use open-source tools in parts of their AI stack, while 76% expect usage to rise. The same research cited lower implementation costs, greater flexibility and easier customisation as major reasons for adoption. Hugging Face has also reported a large and expanding ecosystem, with more than 13 million users, over 2 million public models and more than 500,000 public datasets.
Even so, open source is not automatically the right answer. Quach said Razer would still choose a frontier model when the difference materially affects how an experience is perceived or whether it feels best in class. That distinction matters for products that depend on unusually strong reasoning, premium creative output or a sharply differentiated user experience. Industry leaderboards from groups such as Artificial Analysis continue to show proprietary models occupying many of the top positions in image-quality rankings.
For Korean AI companies, the stakes are particularly high. Start-ups now operate in a market shaped by domestic model developers, local infrastructure investment and rapidly falling access barriers. Upstage’s Solar Pro 2, which KoreaTechDesk reported outperformed GPT-4.1 in the Artificial Analysis Intelligence Index, underlines how quickly the competitive landscape is changing. At the same time, FuriosaAI’s RNGD chip demonstration at the OpenAI Korea launch event pointed to Korea’s growing role in the compute layer, where power and cost remain major bottlenecks.
The broader message is that “good enough” is not a fixed point. It changes as models improve, prices shift and users become more discerning. That makes AI model selection a moving target, especially for startups building education, commerce, gaming, beauty, customer service or companion products. The strongest long-term advantage may belong not to the teams that always choose the most advanced model, but to those that know when a cheaper option is already sufficient and when extra capability truly changes the product.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
