With major models being released more frequently, I’ve been paying more attention not only to benchmark results or model specifications, but to how the differences actually show up in use.
One thing that stands out is that a model becoming cheaper, faster, or more capable on paper does not necessarily mean that its core user experience has improved to the same degree.
For example, when a lightweight model becomes significantly cheaper, that is obviously valuable. But for users who interact with these models every day, the more important question is often much simpler: does it actually feel substantially better at the things they use it for?
If the answer is not obvious, users will naturally compare the product with alternatives based on their actual workflows rather than on the model’s positioning or benchmark performance.
This becomes particularly important when different models develop very different strengths. Coding is a good example. For programmers and other users whose work depends heavily on coding and agentic workflows, small differences in reasoning, context handling, tool use, reliability, and long-running execution can become extremely important. If another model consistently feels more effective for those workflows, users have a very practical reason to move part of their work there.
I don’t think coding should be treated as something OpenAI can simply leave to competitors. Losing a large and highly engaged group of users would be a serious problem, regardless of what the long-term vision is.
At the same time, I don’t think the answer should be to let coding competition define the entire direction of general-purpose AI. OpenAI’s stated goal is much broader than being the best coding model. General-purpose intelligence is supposed to be useful across many different kinds of work and eventually across much more than software.
That creates an interesting strategic distinction. **Some areas require aggressive competition. Other areas require independent direction.** Coding is one of the areas where falling behind can directly cause users to leave. But in the broader definition of general AI, it may be more important to develop strengths that are difficult for competitors to simply copy by targeting the latest release.
This is also where I think edge AI and embodied intelligence become relevant.
I don’t see embodied intelligence as something fundamentally separate from AGI, or as simply “robots with AI added to them.” The physical system is the body; the AI is what gives that system perception, reasoning, adaptation, and decision-making.
From that perspective, **edge AI is a natural extension of general-purpose intelligence into the physical world.** Phones, wearables, cameras, vehicles, industrial robots, drones, medical devices, and other systems all create situations where intelligence needs to be close to the sensors and actuators. Latency, privacy, connectivity, reliability, and real-time decision-making become part of the problem.
And I don’t think this should necessarily be viewed as a distant, speculative opportunity. The important question is not simply whether embodied or edge AI will exist in the future. It is who establishes a meaningful position in that ecosystem early enough to influence its standards, integrations, developer ecosystem, and deployment infrastructure.
The potential impact is enormous because these systems can become part of ordinary production and daily life rather than remaining inside a chat window. That is why I think there is a risk in becoming too absorbed in an endless cycle of reacting to competitors.
A major model is released. Competitors analyze its strengths and weaknesses. The next release targets those weaknesses. The process repeats.
Of course, competition is necessary. **But if every release increasingly becomes a response to the previous release from another company, it becomes possible to lose something more important than a benchmark lead: the ability to decide what problems are worth solving next.** This is also where product quality becomes important.
Users do not experience a model through its benchmark score. They experience it through thousands of interactions. Unnecessary mistakes, redundant responses, unexpected behavior, weak long-context retrieval in smaller models, insufficient usage limits, or situations where a model seems noticeably weaker during an ongoing conversation may each look like relatively small issues. But for someone using an AI system heavily, these small sources of friction accumulate.
Eventually the question is no longer: “Which model is theoretically better?” It becomes: “Which product is easier for me to actually use?”
And that is enough to make people move specific parts of their work to another model, even if they still prefer the original product overall.
I think that is the larger issue worth paying attention to. OpenAI does not need to win every narrow category by copying whoever is currently strongest in it. But it also cannot afford to ignore categories where users are actually moving away.
The more difficult challenge is to do both at once: remain competitive enough that users have no practical reason to leave, while continuing to build toward a broader definition of general-purpose intelligence that competitors are not defining for you.
That broader direction could eventually include not only better cloud models, but agents, edge AI, embodied intelligence, hardware integration, and the ecosystems built around them.
The question is therefore not simply who releases the next better model. It is also who gets to define what “better” means next.