For three years, we asked whether China would catch up with the United States in artificial intelligence. The question assumes a single race with a finish line. That is no longer the case. Depending on whether we're talking about text, video or image, the answer isn't just different: it's sometimes inverted.
We covered this topic in May in our article on the silicon curtain. Three months were enough to make that assessment obsolete. Here is the state of play, with figures to back it up.
Text: America holds the summit, by a very slim margin
This is the most well-documented modality, with independent rankings that are considered authoritative.
| Model | Origin | Intelligence Index | Weights |
|---|---|---|---|
| Claude Opus 5 | 🇺🇸 Anthropic | 63 | Closed |
| Claude Fable 5 | 🇺🇸 Anthropic | 62 | Closed |
| GPT-5.6 Sol / Grok 4.6 | 🇺🇸 OpenAI / xAI | 61 | Closed |
| Kimi K3 | 🇨🇳 Moonshot | 60 | Open |
| Qwen 3.8-Max | 🇨🇳 Alibaba | 58 | Open |
| GLM-5.3-Flash | 🇨🇳 Z.ai | 57 | Open (MIT) |
| Gemini 3.1 Pro | 48 | Closed |
The reading is clear. The top four spots are American, but the gap with the best Chinese model is three points out of sixty-three, or about 5%. Eighteen months ago, that gap was measured in generations.
And the right-hand column tells the other half of the story: China's leading models have their weights published. You can download them, host them, modify them. None of the leading American models allow this.
The most surprising fact in this table isn't Chinese, it's American. Google, long considered part of the top trio, finds itself at 48, far behind. The reason, we've documented it: Gemini 3.5 Pro still hasn't shipped after four delays, and the company only delivers Flash versions. This says something important: the ranking doesn't measure nations, it measures companies. And a company can fall behind.
Video: China ahead, and America retreating
Here, your intuition isn't just right, it's below reality.
In blind rankings, where humans vote without knowing which model produced which clip, the situation is as follows. In text-to-video with audio, Gemini Omni Flash and MiniMax-H3 are separated by three Elo points, which is statistical parity. On another ranking, ByteDance's Seedance 2.0 leads at 1212, with Google's model eight points behind.
In image-to-video, Seedance holds the top spot. And all four variants of Kling, from China's Kuaishou, rank ahead of Veo 3.1, which finds itself in eleventh place.
But the most telling fact is an absence. Sora no longer appears on the rankings, and its API shuts down on September 24, 2026. OpenAI is simply leaving video generation altogether.
In other words: on this modality, the United States now has only one truly competitive player, Google, while China fields four — ByteDance, Alibaba, Kuaishou and MiniMax. And the best open-weights video model is Chinese.
Image: no one wins, and Europe exists
This is the modality where your intuition deserves some nuance. There's no winner, there are specialities.
GPT Image 2 dominates on text embedded in images, which is a major professional use case and long the sector's weak point. Nano Banana 2, from Google, leads on clean output visuals and speed. Two American models, then, at the top on two important axes.
But Seedream, from ByteDance, wins on editorial photography and fashion, and Qwen offers open variants. China isn't absent, it occupies other segments.
And a third player deserves naming: FLUX Kontext, from Germany's Black Forest Labs, is the reference for editing real photographs. It's the only domain where a European player holds a leading position, and we discussed this regarding the release of FLUX 3.
The real dividing line isn't quality
Here's what this overview reveals, and it's more interesting than the medal count.
The difference between the two countries is no longer primarily about capability, which is converging, but about the distribution model.
Leading American labs sell access to closed models, capturing value through usage. Chinese labs massively publish their weights under permissive licences, capturing value elsewhere: in services, infrastructure, or simply in influence.
This strategy has a mechanical effect we've been documenting for months: it exerts constant downward pressure on Western prices. When a model within three points of the top is freely downloadable, it becomes hard to charge a premium for a closed model.
One element changes the analysis. By confirming that the anonymous model Ox Alpha was GLM-5.3-Flash, Z.ai specified that it runs entirely on Chinese AI chips. If this is confirmed, the question of catching up changes in nature: it's no longer about models, but about the fact that the hardware constraint meant to slow China is being circumvented. That was the bet on which export controls rested, and it's the one that will need reassessing.
The caveats, which are important
Three precautions before concluding anything.
Rankings measure what they measure. A composite index aggregates standardised tests that don't necessarily resemble your work. A blind vote on an eight-second clip says nothing about film production. As we keep saying in our article on benchmarks, the only measure that matters is the one you run on your own use cases.
These positions shift every month. This table will be partially wrong in eight weeks. It describes a snapshot, not a trajectory.
Nation isn't the right unit. Google's fall shows this: it's companies that win or lose, with their own trajectories. Talking about China or America as single actors masks considerable internal disparities.
What to take away
The question asked for three years called for a binary answer, and it no longer has one.
On text, the United States retains the summit with a lead reduced to a few percent, and China dominates openness. On video, China leads through the depth of the pack as much as through the scores, while a major American player leaves the field. On image, the distribution is fragmented by use case, with a European player in the leading pack.
What remains true everywhere is that the gap has narrowed to the point where the question of the best model becomes secondary to the question of access conditions: at what price, under what rights, with what ability to host it yourself.
For a European user, that's perhaps the most useful conclusion. The competition between these two countries has produced a result neither of them aimed for: top-tier capabilities, available for free, downloadable and modifiable. Knowing what to do with that is now a more interesting problem than knowing who wins.