我几乎每天都在使用AI大模型，不仅是生活中，更主要在工作中，算的上是大模型的重度使用者了，今天就说下我最近对于大模型的一些体验和感受吧。
I use large AI models almost every day—not only in my daily life, but more importantly in my professional work. It would not be an exaggeration to say that I am a heavy user of large language models. Today, I would like to share some of my recent experiences, impressions, and reflections.
体会
我使用过很多大模型，但现在最常用的还是opus 、Sonnet 、codex，以我个人的体验和实际使用效果来看，我认为对于开发软件来说，这三个目前是最顶级、最聪明的大模型，没有之一。
当然，大模型的发展堪称一日千里，我说的只是当下的情况，以后会如何，who knows?
codex是Open AI的产品，最新的是 codex 5.3 ，claude opus 是anthropic家的，最新的是4.6，我常用的是 claude-4.6-opus-high-thinking 这个。
我自己使用的感受是，codex 5.3 和 claude opus 4.6的水平能力相当，不分伯仲。
但是，codex 在合规方面做了很多限制，而 claude opus显然目前没这方面的约束，对比之下，claude opus用起来明显比codex 更爽。
因此，大多时候，我还是更喜欢用 claude opus，如果硬要说 claude opus有什么缺点的话，我认为它唯一的缺点，就是他妈的太贵了！
就拿今天上午我使用的情况来看，部分费用截图如下，
Experience
I have used many different large models, but the ones I rely on most frequently at present are Opus, Sonnet, and Codex. Based on my personal experience and practical results in software development, I consider these three to be the most advanced and intelligent models currently available for programming tasks.
Of course, the development of large models is progressing at an astonishing pace. My assessment reflects only the current state of affairs. What the future holds—who knows?
Codex is a product of OpenAI, and the latest version I have used is Codex 5.3. Claude Opus is developed by Anthropic, with version 4.6 being the latest. I often use the claude-4.6-opus-high-thinking variant.
From my own experience, Codex 5.3 and Claude Opus 4.6 are comparable in overall capability. Neither clearly outperforms the other in terms of technical proficiency.
However, Codex has implemented a number of compliance-related restrictions. In contrast, Claude Opus currently appears to operate with fewer such constraints. As a result, Claude Opus provides a noticeably smoother and more flexible user experience.
For this reason, I tend to prefer Claude Opus in most situations. If I must point out a drawback, it would be the cost—it is undeniably expensive.
初步统计下，光今天上午我差不多就用了20多美刀，这样的价格，显然最大数人来说不便宜。我自己使用起来也觉得肉疼，因此现在每次使用都是小心翼翼，尽量将问题和内容描述清晰，对比我使用其它的模型那种大手大脚的样子，显然形成了鲜明的对比。囧
For example, based on my usage this morning alone, my expenses exceeded twenty US dollars. Such pricing is far from negligible for most users. Even for me, the cost feels substantial. Consequently, I now approach each session with greater caution, carefully structuring and clarifying my prompts to maximize efficiency. This stands in stark contrast to the more casual manner in which I use other models.
感受
以上都是我最近使用大模型的一些体验，接下来我谈下自己的一点感受和思考吧。
网上和手机上有很多说法，说AI大模型现在这么先进了，要砸很多人的饭碗，包括开发出AI大模型的程序员。
我以前也有过类似的困惑，觉得照这样发展下去，AI大模型取代程序员是迟早的事情，尤其是前阵子，更是有类似的感慨，而通过最近频繁的使用强模型，我这种疑惑感显然逐渐消失了，取代的感受是，不是会取代程序员，而是让程序员这个职业有了更高阶的发展。
诚然，当下的大模型，以及随着肉眼可见的发展速度，AI大模型会取代很多目前的程序员工作，这是大势所趋。但是，这并不意味着要砸程序员的饭碗，而是变相倒逼着程序员这个职业升级，转而升阶到更高段位，比如以前程序员需要逐行编码，经常会为一个bug而消耗几个小时甚至更久时间。
而现在，有了AI大模型，显然可以更快加速这个过程，大大提高效率。换来的是，解放了生产力，让人们有更多的时间花费在更重要和需求的地方，而不是浪费在一些细微末枝、甚至无关痛痒的地方上。
简言之，AI大模型带来的是大大提高效率，大幅度提升人的生产力，让人们有更多时间做更复杂的事情。如同工业革命取代传统手工业，带来的后果不是让社会退步，反而是极大促进了更多生产力，让社会财富更大增加，社会发展进步更快。
Reflections
The above describes my recent experiences. I would now like to share some broader reflections.
There are widespread claims that AI models are becoming so advanced that they will inevitably replace many professions, including software developers—even those who build these models.
I once shared similar concerns. I wondered whether, at this rate of progress, AI would eventually replace programmers altogether. Recently, however, after extensive use of high-performance models, that anxiety has gradually diminished. My current perspective is that AI will not eliminate programmers as a profession; rather, it will elevate and transform the role into a higher-level discipline.
Admittedly, large models—given their visible pace of advancement—will replace certain programming tasks. That trend seems unavoidable. Yet this does not equate to destroying programmers’ livelihoods. Instead, it compels the profession to evolve.
In the past, developers often wrote code line by line and could spend hours, even days, debugging a single issue. Today, with the assistance of large AI models, these processes can be significantly accelerated. Productivity increases dramatically.
What we gain is not redundancy, but liberated capacity—time and cognitive resources that can be redirected toward more meaningful and complex challenges, rather than being consumed by repetitive or low-level tasks.
In essence, large AI models dramatically enhance efficiency and human productivity. They allow individuals to focus on higher-order problems. This transformation resembles the Industrial Revolution: traditional handicrafts were replaced, yet society did not regress. On the contrary, productivity expanded, wealth increased, and social development accelerated.
思考
最后，在这篇文章的末尾，我也想谈下自己的一点忧虑，未雨绸缪，也可能是杞人忧天吧。
还是以我个人的体会，随着使用AI大模型的频繁和深入，我越来越感受到离不开AI大模型，我们都知道当下是AI高速发展和竞争时代，无论国外还是中国，各类大模型都在蓬勃发展，每天都有推陈出新，全世界这么多大模型种类，有种逐鹿中原，不知最后鹿死谁手的感觉。
当下，正因为是竞争时期，每家大模型为了更好的吸引和拉拢用户，价格都很便宜甚至不惜免费，有种赔本赚吆喝的感觉，这对于消费者的我们自然是好事。
但是，任何事情都不可能一直这样免费下去，天下始终没有免费的宴席，如果有，也不可能一直免费下去。大模型亦如此，逃不脱历史发展的必然。想一想当年网约车开始的时候，想一想外卖大战，资本在开始的时候，需要用户和流量，需要培养用户和市场，可以不惜一切砸资金，但是这样的日子总是会过去，等到大模型大战趋于结束，市场垄断开始形成时，就是资本开始收割的时候了。
试问，到那个时候，作为消费者的我们，已经习惯了大模型给我们带来的便利，离不开AI大模型时，而定价权掌握在极少数几家垄断企业手里时，人为刀俎我为鱼肉，试问真到了那个时候，我们还有别的选择吗？
Concerns
Finally, I would like to express a concern—perhaps prudent foresight, perhaps unnecessary worry.
As my reliance on large AI models deepens, I increasingly sense how indispensable they have become. We are currently in a period of intense global competition in AI development. Both internationally and within China, numerous models are emerging rapidly. New advancements appear almost daily. The landscape feels like a modern contest for dominance, and it remains uncertain which players will ultimately prevail.
At present, competition keeps prices low. In some cases, services are even offered at a loss to attract users and build market share. From a consumer perspective, this is beneficial.
However, no business model can remain indefinitely subsidized. History suggests otherwise. Consider the early days of ride-hailing services or food delivery platforms: companies invested heavily to capture users and shape market habits. But once competition stabilized and consolidation occurred, pricing power shifted.
If a similar consolidation happens in the AI industry, and market dominance concentrates in the hands of a few major providers, pricing authority will follow. By that time, if users have grown deeply dependent on these tools, their bargaining power may be limited.
When convenience becomes necessity, and alternatives diminish, what choices will remain?