Just past 1 a.m. today, with England still trailing the Democratic Republic of the Congo after more than 60 minutes into the second half, our World Cup viewing group began earnestly discussing whether "Tuchel should be sacked" and "how British football hooligans might stir up trouble." Then, like a deus ex machina, Kane scored a goal. We started contemplating whether "the match could conclude within 90 minutes without extra time." Football-savvy friends said it was unlikely, as the Democratic Republic of the Congo's defense was robust and their determination unwavering. It seemed the match would have to be decided in 120 minutes.

Then, Kane descended from the heavens once more, and England narrowly advanced. Coincidentally, Gordon, who had just come on as a substitute both times, provided the assists. Tuchel went from the brink of being sacked to the brink of deification, akin to a stock suddenly surging from a limit down to a limit up. However, to be fair, the deciding factor in the match was straightforward: the individual prowess of a superstar.

The same principle applied to Mbappé's double strike the day before yesterday and Haaland's goal. Strictly speaking, in the match between Brazil and Japan, Vinícius Júnior played a similar role: although he didn't score, every time he charged into Japan's defensive third, he caused chaos, fully demonstrating the dominance of individual ability. Japan lacked superstars and had few top-tier players. Even if they had held on into extra time, their chances of winning would have been slim—this should be acknowledged.

This is the World Cup knockout stage, where a single match determines victory or defeat. After 90-120 minutes, one team advances while the other goes home. In such situations, team spirit alone is insufficient; there must be a moment of brilliance from a superstar. It reminds me of Moneyball, where the protagonist relies on data analysis to excel in the regular season but consistently fails to secure the ultimate championship in the playoffs. The movie itself provides the answer:

1. The best players come with a premium. Relying on data analysis to find cost-effective players means that none of your positions are truly top-tier. There is a chasm between excellence and the pinnacle.

2. In knockout matches, you must have the ability to defeat any opponent. At critical moments, you need a star player to "solve the problem." This "problem-solving" ability is expensive, and there are no cost-effective alternatives.

I really like the documentary The Last Dance, which devotes an entire episode to describing Game 6 of the 1998 NBA Finals—the game where Jordan became a god. Pippen was injured at the start, and the home team, the Jazz, were highly motivated. Jordan was physically drained in the second half, missing several shots and only able to score through drives and free throws. In the final moments, Jordan rose up in the same area where he had missed several times before and sank the last shot, securing his sixth championship.

The Jazz lost. Was it fair? No. That's why Jordan was worth $30 million a year, and everyone had to build the team around him. In reality, Malone and Stockton were great players, but there is indeed a gap between mortals and the god of basketball. Unreasonable problem-solving ability is the biggest gap.

After discussing competitive sports at length, what does this have to do with large models? A lot. Recently, I've seen several articles on my social media feed condemning Anthropic, even calling it an "evil force."

The world has long suffered from Anthropic: its pricing is exorbitant, its customer service is poor, it's unfriendly to users from certain countries, and it doesn't allow access to its new models, etc. The more you use Claude, the more frustrated you become with its developer, eagerly awaiting someone to replace it.

DeepSeek V4, Kimi K2.5, GLM 5.2... were all hailed as "Claude killers" by the AI developer community upon their release. This not only shows that they have certain strengths but also highlights how unpopular Anthropic has become. Even DeepSeek, after its price increase, has an API price that is only a fraction of Claude's. As an ordinary Claude Pro user, my biggest pain is that the model quota runs out too quickly. Even if you don't program and only do text analysis, it's like rain in the desert—you barely see a shadow before it's gone.

If Claude weren't so powerful, I doubt anyone in the world would use it. That's both a true statement and a trivial one. We could also rephrase it like this:

I'm a heavy user of GPT with an uninterrupted paid history of 27 months; even I started defecting to Claude two months ago (though I still keep GPT Plus). Because some tasks can only be completed by Claude, or more precisely, only by Claude Opus 4.7/4.8 (plus the newly restored Fable 5). GPT falls just short, and as for other models, I don't even want to talk about them.

Apart from its widely recognized programming and agent capabilities, Claude also provides the best answers for semantic analysis of complex long texts, such as novels, screenplays, philosophical works, and non-fiction. There was a time when GPT was the best (especially during the GPT 4o era), but it has been getting worse in the past six months. It's not that it has regressed; rather, it can't keep up with Claude's pace of improvement. Only Claude Opus can detect certain subtle meanings. I'd also like to point out that Opus is very accurate in identifying "intertextuality," i.e., the mutual referencing, adaptation, and parody between texts. Its accuracy approaches that of a doctoral student in related humanities fields, and it only takes a few seconds to see these things, not days.

Can cheaper models be used? Yes.

Are cheaper models useful? Yes.

Are cheaper models convenient? Yes, at least they won't act like Anthropic, holding a big stick and looking at everyone with disdain.

Can cheaper models complete tasks? Well, that's a more complex question. It depends on the task...

After Lewandowski left, Choupo-Moting once shone for Bayern. Why did Bayern still buy Kane? Was Choupo-Moting useless? Or was he not cheap or cost-effective enough?

Arsenal spent £100 million on Rice. Does that mean there are no cost-effective defensive midfielders in the world? This summer, it seemed like all the Premier League giants were vying for Anderson at £100 million. Have they all lost their minds? Are defensive midfielders worth £30-50 million not good enough?

And what about Olise, who had two good seasons at Bayern? Why did Real Madrid suddenly offer €200 million for him? If Bayern were willing to sell, I believe more than just Real Madrid would be willing to buy at a similar price.

The answer has never been on the data analysis sheets; it's on the field. Kane's two goals this morning, Olise's two assists yesterday, and Rice and Anderson's ability to cover the entire field up and down prove that they are stars with problem-solving abilities at critical moments. Don't forget that in this year's Champions League, Rice scored two free kicks in a single match against Real Madrid. Is this unreasonable problem-solving ability worth £100 million?

So even if Anthropic is as evil and arrogant as it is, everyone has to put up with it. As soon as Claude Fable 5 was restored this morning, I immediately burned through a wave of model quota. Unless GPT-5.6 is fully released and significantly improved, I think Claude will remain my most important productivity + life assistant tool this year. Since I don't raise lobsters and don't have many daily agentic workflow tasks, cheaper large model APIs are not very useful to me. I'll still have to painfully give Anthropic my money.

By the way, after losing trust in Gemini for more than a month, I became a paid user of Google AI again. The sole reason was Nano Banana Pro's image generation capabilities. I compared GPT Image-2 and Nano Banana Pro and believe that the former is powerful in many ways but still falls slightly short of the latter. I want to generate character designs and scene images in Chinese or Japanese anime styles. For some reason, GPT is relatively poor at this. Almost all professional AI anime companies use Nano Banana Pro as their main image generation model, not because other models are useless, but because none of them can consistently meet the highest standards.

Now, the most embarrassing situation is for OpenAI. It is said that Sam Altman decided to delay the IPO because it would be difficult to guarantee a valuation of over $1 trillion. Then the question arises: Anthropic's valuation in its last funding round was already $962 billion, just a step away from $1 trillion. Why couldn't OpenAI, which reached an $860 billion valuation earlier this year, take it a step further?

Because GPT is no longer the "best model," just "one of the best models." It is slightly inferior to Claude in text tasks and coding and clearly cannot match Gemini (and Seedance) in multimodal capabilities. This is like a player who is not positioned as a "superstar" and lacks unconditional problem-solving ability, but is merely a "qualified first-tier player"—the actual gap may only be 5-10%, but it's the difference between heaven and earth.

Because the strongest model holds pricing power, it is destined to capture the majority of the market's profits. Competitors can take a cost-effective route, but strictly speaking, this route does not compete with the "strongest model," just as rotational players on a team do not compete with superstars. To earn a superstar's salary, there is only one way: reach superstar level in terms of ability, i.e., possess problem-solving ability.

It seems that everyone dislikes Mbappé, but he can still score at the World Cup, and Olise can still feed him the ball. That's why these two are at the top of the superstar valuation list. One day, when Mbappé can no longer score or scoring doesn't change his fate, he'll come down from the pedestal without anyone needing to criticize him. There's a bit of that at Real Madrid, but it's still early.

The bottom line is: even if Mbappé comes down from the pedestal, his replacement will surely be another superstar forward worth €100 million or even €200 million, not the sum of several €30 million forwards. Real Madrid's use of Josélu and Bayern's use of Choupo-Moting are temporary measures; otherwise, the fans would revolt, and the team couldn't maintain the highest level of competitiveness.