Table of Contents
- The 3% Gap Itself
- The 2.7% Stanford Number
- The 39-Point Arena Difference
- The 6% Reading From June
- The 23x Money Gap
- Gemini 4 Argon’s 1 Million Token Output
- Argon’s $2 and $10 Price Tag
- DeepSeek’s Permanent 75% Discount
- The $0.28 Floor
- The 1,100 Models Problem
- FAQ and Final Verdict
Introduction
Three percent is the kind of number most people scroll past. A restaurant tip. A small rate change. Nothing worth stopping for.
Then you learn it is the reported benchmark gap between China’s best AI models and the best American ones, and it gets harder to scroll.
Let me be clear about what this post is not. It is not a victory lap for China, and it is not a panic button for the US. Benchmarks measure test scores. They do not measure who has more chips, more money or more customers. And as you will see, even the 3% figure changes depending on who counts it and when.
What I did instead was collect 10 numbers that explain where things really stand on October 20, 2026. Some come from Stanford and Bloomberg Intelligence. Some come from the launch of Google’s new Gemini 4 Argon. Some come from the DeepSeek-led price war that is making AI cheaper by the month.
Every number below has a source behind it, and I will tell you where a figure is solid and where it is shaky. Let’s start with the one in the title.
If you want the political side of this race first, my breakdown of Trump’s AI race with China shows how Washington has been reacting to these gap numbers.
Number 1: The 3% Gap Itself
The claim: Chinese models are about 3% behind leading US models on benchmark scores.
This comes from Bloomberg Intelligence, as reported by YourStory in a piece that credits DeepSeek’s newer models with helping close the distance. The same article adds a useful reality check. Only three of the top 15 models on LiveBench are Chinese, and it stresses that a 3% gap does not mean China has overtaken the US.
What 3% really means?
Think of two sprinters in a 10 second race. A 3% gap is 0.3 seconds. That is small enough to feel like a tie, and big enough to decide gold versus fifth place.
That is exactly where Chinese models sit. They are close enough that you can use them for many daily tasks and barely notice a difference. They are still not the ones standing on top of the podium.
The catch
I could not confirm the exact publication date of this Bloomberg Intelligence reading. Treat it as the latest figure available rather than a measurement tied to one specific day. If you quote it anywhere, say “according to Bloomberg Intelligence” and name the benchmark approach, not just the number.
Number 2: The 2.7% Stanford Number
The claim: The top US model led its nearest Chinese rival by 2.7% as of March 2026.
This one comes from Stanford’s 2026 AI Index, one of the most cited yearly scorecards in the field. Stanford’s own technical performance chapter is where the data lives, and several outlets picked up the 2.7% headline after the April release.
Why it sits so close to the 3%?
Both numbers tell the same story from different angles. Stanford looked at the Arena leader-board in March. Bloomberg Intelligence tracks benchmark scores month by month. Two teams, two methods, and both land in the same neighbourhood.
That agreement is the strongest evidence in this whole post. When independent scorekeepers reach similar results, the claim is much harder to dismiss.
A fair warning
The Index itself notes that Arena style rankings can be gamed. Implicator’s breakdown of the report points out that extra interaction data can help a model climb the ladder. So a small gap on one leader-board is a strong hint, not a final verdict.
Number 3: The 39-Point Arena Difference
The claim: Top US model 1,503 points, top Chinese model 1,464 points.
Here is the raw score behind the 2.7%. In March 2026, The Next Web reported that Anthropic’s Claude Opus 4.6 led the global Arena leaderboard at 1,503, while ByteDance’s Dola-Seed-2.0-Preview sat at 1,464. Divide 39 by 1,503 and you get roughly 2.6%, which matches Stanford’s rounded headline.
Look at how far it fell
In May 2023, the story was completely different. Fortune’s coverage of the report notes that GPT-4 led with more than 1,300 Arena points while the best Chinese model had fewer than 1,000. A gap of several hundred points shrank to 39 in under three years.
That drop is the real headline. The 3% is the snapshot. The collapse from hundreds of points to dozens is the trend.
The model sitting at the top of that list has since been replaced by newer releases, and my Claude Opus 5.5 vs GPT-6 Sol vs Grok 4.7 comparison shows how the leaders compare today.
Number 4: The 6% Reading From June
The claim: The gap hit a record low of 6% in June, down from 9% in May.
Wait, is that not bigger than 3%? Yes, and this is where most blog posts get sloppy, so let me slow down.
The Next Web reported that Bloomberg Intelligence measured a 6% gap in June, which it called a record low at that time. Over the previous 12 months, the same research arm said the gap averaged between 10% and 15%. Zhipu AI’s GLM-5.2 reaching the top of a global agentic coding ranking helped push the number down.
How the gap moved?
| Reading | Gap | Source | Notes |
|---|---|---|---|
| Prior 12 month average | 10% to 15% | Bloomberg Intelligence | Before mid-2026 |
| May 2026 | 9% | Bloomberg Intelligence | Month before the drop |
| June 2026 | 6% | Bloomberg Intelligence | Record low at the time |
| March 2026 | 2.7% | Stanford AI Index | Arena leaderboard method |
| Latest reading | About 3% | Bloomberg Intelligence | Benchmark scores, exact date unconfirmed |
What the table tells you?
Different methods produce different numbers, and the months matter. The honest summary is not “the gap is exactly 3%.” The honest summary is that the gap has been shrinking fast and now sits somewhere between roughly 3% and 6%, depending on the benchmark you pick.
That wording will hold up even when the next report lands, which makes it safer for you and for your readers.
Number 5: The 23x Money Gap
The claim: US private AI investment hit $285.9 billion in 2025, about 23 times China’s $12.4 billion.
Here is the number that makes the 3% gap look strange. If money alone decided the race, China would not be anywhere near the US. Stanford’s data, summarized by AI Weekly, puts American private AI spending at roughly 23 times the Chinese figure, and says the US hosts 5,427 data centers, more than 10 times any other country.
So how does a country spending far less end up this close? A few reasons stand out.
Why the gap did not follow the money?
- Efficiency beat size. Chinese labs shipped open models and cheaper designs instead of trying to outspend everyone.
- Other strengths count too. Stanford’s figures, as covered by Logos Press, show China filing about 69.7% of global AI patent applications and installing industrial robots at a far higher rate than the US.
- The US talent pipeline is thinning. The same AI Weekly summary notes that AI researchers relocating to the US have fallen 89% since 2017, with 80% of that drop in the last year alone.
The fair counterpoint
Private investment is one slice of the pie. It does not capture government spending, and it does not capture who owns the chips. Money still buys the most expensive training runs, and the US still leads in the number of top models. The 23x figure explains why the US is still ahead. It does not explain why the lead is so small.
Spending that big also raises the question of whether the whole thing is sustainable, which I looked at in my piece on the AI bubble numbers behind OpenAI, Anthropic and SpaceXAI if you want the financial side.
Number 6: Gemini 4 Argon’s 1 Million Token Output
The claim: Google says Gemini 4 Argon can generate up to 1 million tokens in a single run.
Now we move from the gap itself to the American side of the race. Google’s newest flagship is the model that most people will compare Chinese releases against for the next few months.
The timeline is worth knowing. Sundar Pichai gave an early look on September 30, and the model reached paying users and developers on October 13, when Google announced the wider release of Gemini 4 Argon. Early access went first to the US government and a group of cyber defense partners.
What 1 million tokens of output really means?
A token is a small piece of text, roughly three quarters of a word. One million tokens is somewhere around 750,000 words in a single answer. That is more than the entire Lord of the Rings trilogy.
Nobody needs that in a chat reply. The point is long jobs. HPCwire’s launch coverage describes it as a big jump for long code migrations and multi-step agent work, and says Argon uses a sparse mixture-of-experts design that sends each input to a small group of specialized sub-networks.
Read the fine print
Google’s own developer documentation for Argon shows the default output cap is 65,536 tokens, and the full 1,048,576 only unlocks with a long decode setting. So the headline figure is real, but you have to switch it on.
Where this meets the China story?
Argon arrived after most of the gap readings in this post. I have not found a source that measures the 3% against Argon, so I will not pretend one exists. What I can say is that the next gap reading will be the first test of whether a fresh American flagship pushes the number back up.
Google has been pushing hard against its biggest rival lately, and my look at the Google vs OpenAI AI war explains why this launch matters beyond the benchmarks.
Number 7: Argon’s $2 and $10 Price Tag
The claim: Argon costs $2 per million input tokens and $10 per million output tokens through November 30.
This is the number that connects Google to the price war. The launch pricing is a limited-time deal, and the normal rate is double.
| Price type | Input (per 1M tokens) | Output (per 1M tokens) | Source note |
|---|---|---|---|
| Launch promo, until Nov 30 | $2 | $10 | Reported by HPCwire |
| Standard rate | $4 | $20 | Reported by iThome |
| Cached input | About 95% cheaper than normal input | Not applicable | Per iThome |
The standard rates and the cached input detail come from iThome’s report on the Argon rollout, which also notes the promo ends on November 30, 2026.
A quick example
Say you use the full 1 million token output limit once. At promo pricing, that output alone costs $10. At the standard rate it costs $20. Add the input you send in, and a giant run still costs less than a lunch.
That sounds cheap until you run thousands of them. Companies using agents run jobs in bulk, and this is exactly where Chinese competitors are pushing hard. Which brings us to the real fight.
Why this matters for the gap?
Google can charge premium prices when it leads. The fact that it is discounting at launch, with a clear end date, tells you the market is competitive. In the next part, I will show you how much cheaper the Chinese side has gone.
Number 8: DeepSeek’s Permanent 75% Discount
The claim: DeepSeek made a 75% price cut on its flagship V4 Pro model permanent.
This is where the price war stops being a headline and becomes a business model. It started as a launch promo. Bloomberg reported in April that DeepSeek was pitching a 75% discount on V4-Pro to developers the week after release, and had cut input cache hit fees to a tenth of the old price across its models.
Promos come and go. What changed the mood was the next step. The Next Web reported that DeepSeek made the discount permanent, and that V4 Pro keeps a one million token context window at the new price. In that report, the permanent rate sits below Google’s budget Gemini 3.5 Flash, which charges $0.15 for input and $0.60 for output per million tokens.
What “permanent” does to the market?
A temporary sale makes buyers wait. A permanent price makes buyers plan. Teams that run thousands of agent tasks a day can build a budget around a number that is not going to jump next month.
One newsletter put the new V4 Pro rate at $0.27 input and $0.87 output per million tokens. I could not confirm that against DeepSeek’s own pricing page, so check the official page before you quote it anywhere.
The risk side
The same report raises a fair question for business buyers. Cheaper is not always better when you factor in reliability, compliance and the risk of sending sensitive work through a foreign provider. Price wins the first meeting. Trust wins the contract.
Cheap open models are also changing who can build with AI at all, as I explained in my guide to Mistral’s Le Chonk and its open weights for teams weighing self-hosted options.
Number 9: The $0.28 Floor
The claim: DeepSeek’s V4-Flash-0731 prices output at $0.28 per million tokens.
If $0.87 sounds cheap, this next one is almost a rounding error. Forkast reported that DeepSeek released V4-Flash-0731 on July 31, 2026, priced at $0.14 for input and $0.28 for output per million tokens. The timing mattered too. It landed two days after OpenAI cut prices by 80% on its GPT-5.6 Luna model, so it read like a direct counterpunch.
The same report says DeepSeek’s changelog shows the model scoring 82.7 on Terminal Bench and 76.7 on Cybergym, which are tests aimed at agent style work.
Compare it side by side
| Model | Input per 1M tokens | Output per 1M tokens | Note |
|---|---|---|---|
| Gemini 4 Argon (standard) | $4 | $20 | Launch promo is $2 and $10 |
| Gemini 3.5 Flash | $0.15 | $0.60 | Google’s budget model |
| DeepSeek V4-Flash-0731 | $0.14 | $0.28 | Reported at launch, July 31 |
Look at that bottom row against the top row. Output on the Chinese budget model costs about 70 times less than Argon at standard pricing. Different tiers, different jobs, but you can see why buyers started paying attention.
The plot twist nobody expected
Here is where I have to be honest with you, because a lot of price war posts skip this part. Just days after those launches, DeepSeek warned of a significant API price increase, without sharing new rates, affected models or a timeline. The same analysis notes that Moonshot released its Kimi K3 open-weight model and then paused new consumer signups, which points to demand running hotter than capacity.
Zhipu AI moved in the same direction. According to the South China Morning Post, it raised prices while open sourcing its flagship model, and now charges $1.40 per million input tokens and $4.40 per million output tokens.
So the “new AI price war” has two halves. Prices fell hard in the first half of 2026. Then some Chinese labs started edging them back up. The floor is real, but it may not be permanent for every model.
The smarter way to read prices
The VKTR piece points to an Artificial Analysis finding that a model costing more per token can still be cheaper in practice if it finishes work in fewer turns. In plain words, the sticker price is not the bill. The total cost of finishing a job is.
Number 10: The 1,100 Models Problem
The claim: China has more than 1,100 language models fighting over the same customers.
The same Bloomberg Intelligence reading behind the 3% figure adds a detail that changes how you should see the whole story. China has more than 1,100 language models, and that crowd is a big reason a price war over tokens broke out in the first place.
When over a thousand products compete for the same buyers, the only easy lever is price. That is great for customers right now. It is painful for the companies, because the same report points out that China still has a profitability problem.
Two races at once
Chinese labs are running two races at the same time:
- Race one: close the performance gap with the best American models.
- Race two: find a way to make money while prices keep falling.
The US labs face a version of this too, but with deeper pockets, as the 23x spending gap from Number 5 shows. That money cushions them while prices drop.
One more reason to stay careful
A fresh example of how fragile any single gap number can be comes from an analysis of Moonshot’s Kimi-K3, a 2.8 trillion parameter open-weight model announced on July 16, 2026. The author of that Substack breakdown argues that benchmark choice can change the answer about how far behind China really is, and says Kimi-K3’s coding ability looks comparable to current US frontier models.
That fits the pattern of this whole post. On coding, the gap can look tiny. On broad leaderboards, it can look bigger. Anyone who tells you the gap is one fixed number is selling something.
What the 10 Numbers Say When You Put Them Together?
Ten numbers can feel like a pile. Here they are in one place, so you can see the shape of the story.
| # | Number | What it tells you | How solid is it? |
|---|---|---|---|
| 1 | 3% | Benchmark gap, latest reading | Strong source, date unconfirmed |
| 2 | 2.7% | Stanford’s March 2026 gap | Strong |
| 3 | 39 points | 1,503 vs 1,464 on Arena | Strong |
| 4 | 6% | Record low in June, before the latest reading | Strong |
| 5 | 23x | US vs China private AI spending | Strong |
| 6 | 1M tokens | Argon’s output ceiling | Strong, but off by default |
| 7 | $2 and $10 | Argon’s promo price until Nov 30 | Good, standard rate from one outlet |
| 8 | 75% | DeepSeek V4 Pro cut made permanent | Good |
| 9 | $0.28 | DeepSeek budget model output price | Medium, one outlet |
| 10 | 1,100+ | Chinese language models competing | Medium |
The pattern is simple. The performance gap is small and shrinking. The money gap is huge. The price gap is the weapon Chinese labs use, and American labs are answering with discounts of their own.
Who Actually Wins From This Race?
Developers and small businesses win first.
When top level models cost a few dollars per million tokens, or less than one dollar for budget tiers, projects that were too expensive last year suddenly work on a small budget.
US labs win on trust and the top end.
Stanford’s data shows the US still leads in total top models and in the hardware that trains them. And when the work is sensitive, many buyers will still choose a US provider even if it costs more.
Chinese labs win on speed and openness.
Open weights and low prices put their models into places a locked, expensive model never reaches. The catch, as Number 10 showed, is that making money is still the hard part.
The race is also shaping how governments talk about AI itself, which I covered in my post on Trump’s order to say “Super Intelligence” instead of AI and what the 120 day task force is meant to do.
Three Things Still Unproven
Good reporting means saying what we do not know. Three claims floating around the internet are not backed up yet.
1. That China has overtaken the US.
The State of AI report’s tracker says Chinese models have reached the top five on LMArena’s overall leaderboard, but none has taken and held first place. Close is not the same as ahead.
2. That the gap is only about chips.
A KrASIA report on CITIC CLSA’s view says China’s gap has narrowed to about three months thanks to fast iteration, open source and chip reserves, but that advanced chip production remains a longer term hurdle. Software progress can hide a hardware problem for a while.
3. That the gap will keep shrinking in a straight line.
Fresh American releases like Argon can move the number back the other way. Last year, a Recorded Future analysis put the lag at three to six months, and said the time lag was shortening. A trend is not a guarantee.
My look at Musk’s claim that SpaceXAI will lead the AI race within six months shows how quickly timelines like these get tested.
FAQ
Is China’s AI really only 3% behind the US?
On benchmark scores, that is what Bloomberg Intelligence reported. Stanford found 2.7% in March, and Bloomberg Intelligence found 6% in June. The number depends on the benchmark and the month, so “roughly 3% to 6%” is the safest way to say it.
Has a Chinese model reached number one on a major leader-board?
Not in a lasting way, according to the State of AI tracker linked above. Chinese models have made the top five, but none has held the top spot.
Is Gemini 4 Argon open to everyone?
It reached paying Gemini plans and developers on October 13. The newest app features were announced for Pro and Ultra subscribers, so free users should not expect it yet. Check Google’s page for the latest access details.
Will DeepSeek prices stay this low?
V4 Pro’s 75% cut was made permanent, but DeepSeek also warned in August of a significant API price increase without sharing details. Always check DeepSeek’s official pricing page before you plan a budget around any number in this post.
Does a 3% benchmark gap mean the models feel the same to use?
Not always. Benchmarks measure test performance. Reliability, safety, speed and how well a model fits your task can feel very different in daily use.
Final Verdict
So, is China’s AI only 3% behind the US? On benchmarks, close to it. That is a real and remarkable result for a country spending about 23 times less on private AI investment.
But a small gap on a test is not the same as parity everywhere. The US still leads at the very top, in money and in hardware. China leads in price, openness and patents. And the price war that makes this fun for buyers is also squeezing the labs behind it.
My honest read is this. The era of a comfortable American lead is over. The era of a clear Chinese winner has not started. What you have is a close, fast, expensive and cheap race all at once, and the next gap number will probably look different from this one.
If you build with AI, test more than one model and track the total cost of finishing a job, not just the sticker price.