Will Chinese open models beat OpenAI?

Why is this question serious now?

Chinese open models are a serious threat because the pressure no longer depends on one lab or one release.

DeepSeek first showed that a Chinese laboratory could build strong reasoning models under tight hardware limits. Since then, Z.ai, Moonshot AI, Alibaba’s Qwen team and MiniMax have kept pace, and newer entrants have joined them. In October 2026, Xiaomi’s MiMo-V2.6-Pro led Artificial Analysis’s open-weight ranking, and StepFun’s Step 5 Preview was in the top five.

Usage has followed. On OpenRouter, the largest marketplace for model APIs, Chinese labs handled more requests than OpenAI in mid-October 2026. Seven of the ten most-used models by tokens were Chinese.

Revenue has followed too, from a small base. Z.ai’s first-half 2026 revenue was about five times the year before. MiniMax’s rose 283.1%. DeepSeek’s annualized revenue passed $1 billion in September, according to The Information.

What has changed is the target. OpenAI is no longer the clear leader Chinese labs are chasing. Anthropic, Meta and Google now share or hold the top of the main leaderboards. Chinese open models put pressure on OpenAI from below while American rivals press it from above.

What would “beating OpenAI” actually mean?

Chinese open models can beat OpenAI’s models without replacing ChatGPT or overtaking OpenAI’s business.

A technical victory would mean similar or better intelligence on the tasks users care about. Such a lead can vanish with the next release, so one leaderboard result is not a permanent win.

An ecosystem victory would last longer. Chinese models would become the models developers download, modify, host and build into their own products, with tools and expertise growing around them.

Beating OpenAI overall requires much more. OpenAI reported more than 1.2 billion weekly ChatGPT users on October 7, 2026. It had more than 50 million paying consumer subscribers in March 2026, its latest disclosed count, and said in September 2026 that it serves 2.5 million businesses.

For this article, a model-layer victory means Chinese open models combine near-frontier quality with lower prices and more customer control. An outright victory would require those advantages to produce stronger global distribution, products and revenue than OpenAI.

By October 2026, the first outcome is partly achieved: Chinese models lead on price-for-capability and on marketplace usage, but not on raw capability. The second remains distant.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

How far behind OpenAI are Chinese open models now?

About eight points on the main independent index, which is close enough for many buyers but not close enough to call them equal.

On October 20, 2026, Artificial Analysis gave GPT-6 Astra 53 on its Intelligence Index. The best Chinese open models scored 44 to 46: Xiaomi’s MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash and Z.ai’s GLM-5.3 45, and Moonshot’s Kimi K3 44.

Artificial Analysis now uses version 4.3.2 of its index. In July 2026, on the previous version, Kimi K3 scored 57 against 59 for GPT-5.6 Sol, but scores from the two versions cannot be compared. On the current version, the Chinese gap looks wider.

Two other facts matter more than the exact gap. First, OpenAI is no longer at the top: Claude Opus 5.5 scores 58 and Meta’s Muse Wonder Full 56. Second, Kimi K3 is now open weight. Moonshot released its weights in July 2026 under a custom license, so it counts as an open model, with conditions.

Arena’s human-preference rankings tell a similar story. On October 20, 2026, Google, Meta and Anthropic held the top three places for overall text. Kimi K3 was the best Chinese model, in 18th place.

Most businesses do not buy the top score. They buy a model that handles their task reliably at a reasonable cost. Chinese open models are in that range for coding, extraction and agent work. They are not yet at the frontier.

Model	Availability	Intelligence Index	Gap from GPT-6 Astra
Claude Opus 5.5 (Anthropic)	Closed API	58	+5
Muse Wonder Full (Meta)	Proprietary	56	+3
GPT-6 Astra (OpenAI)	Closed API	53	-
MiMo-V2.6-Pro (Xiaomi)	Open weight	46	-7
GLM-5.3 (Z.ai)	Open weight	45	-8
DeepSeek V4.1 Flash	Open weight	45	-8
Kimi K3 (Moonshot AI)	Open weight, custom license	44	-9
Muse Spark Open (Meta)	Open weight, custom license	39	-14
GPT-6 Luna (OpenAI)	Closed API	38	-15
OpenAI gpt-oss-120b	Open weight	12	-41

Is OpenAI still the frontier Chinese labs are chasing?

No longer alone. In 2026, the labs pushing the frontier forward were mostly Anthropic, Meta and Google, and Chinese labs are chasing all of them.

The race is still a cycle. An American lab releases a stronger model. Chinese labs approach that level faster and more cheaply than before. The gap shrinks until the next release.

The cycle sped up in the fall of 2026. Anthropic released Claude Opus 5.5 on September 22. OpenAI released GPT-6.1 Sol and Luna the same day and rolled out GPT-6 more broadly on October 7. Meta released Muse Wonder on October 12, and Google made Gemini 4 Argon broadly available on October 13.

Chinese labs kept shipping. DeepSeek released V4.1 Flash on September 10 and a revised version in October. Z.ai moved to GLM-5.3. MiniMax released M3.1 Flash on October 19. None reclaimed a frontier position, but each stayed in the leading open group.

The American labs hold a compute advantage that is growing. Meta’s Prometheus cluster, which trained Muse Wonder, is being expanded toward about one million GPUs. Epoch AI estimates xAI’s Colossus 2 at about 1.1 million H100-equivalents.

Epoch AI estimated in January 2026 that Chinese models had trailed US models by seven months on average, within a range of four to 14 months. Nothing in the October rankings suggests that lag has closed. For OpenAI, the bigger change is that the frontier is now contested by three American rivals.

Can OpenAI’s cheaper GPT-6 models match Chinese prices?

On list price, OpenAI now matches Chinese labs at the low end. At comparable capability, Chinese models remain several times cheaper.

OpenAI changed its price ladder on September 22, 2026. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. That is below DeepSeek’s off-peak rate for V4.1 Flash, $0.15 and $0.60. GPT-6.1 Sol costs $2 and $10, and GPT-6 Astra, OpenAI’s strongest model, $10 and $50.

The price per unit of capability tells a different story. Luna scores 38 on Artificial Analysis’s index. DeepSeek V4.1 Flash scores 45, and GLM-5.3 45. To beat those scores with OpenAI, a buyer needs GPT-6.1 Sol, which scores 52 and costs about 13 times more than off-peak DeepSeek for input and 17 times more for output.

DeepSeek’s prices also depend on the clock. Peak rates, on weekday mornings in UTC, are double the off-peak rates. Z.ai charges $1.40 and $4.40 for GLM-5.3, and $0.15 and $0.50 for GLM-5.3-Flash.

Google is squeezing the premium tier from the other side. Gemini 4 Argon’s introductory price is $2 and $10, the same as GPT-6.1 Sol, for a model that scores 53. Google’s launch terms say the price rises to $4 and $20 after the introductory period.

The engineering behind Chinese prices is real. DeepSeek V4.1 Flash has 552 billion parameters but activates only 8 billion on input and 16 billion on output. Public prices do not prove that every endpoint is profitable, but subsidies alone cannot explain the gap.

The result is a narrower but still clear advantage. OpenAI can match Chinese prices on simple tasks. For mid-level coding and agent work, Chinese open models still deliver more capability per dollar.

Model	Input per 1M tokens	Output per 1M tokens	Overall position
GPT-6 Astra	$10.00	$50.00	OpenAI’s strongest model, index 53
GPT-6.1 Sol	$2.00	$10.00	Mid-priced OpenAI model, index 52
Gemini 4 Argon	$2.00	$10.00	Google, introductory price, index 53
GLM-5.3	$1.40	$4.40	Z.ai open model, index 45
DeepSeek V4.1 Flash (off-peak)	$0.15	$0.60	Open model, index 45; double at peak
GPT-6 Luna	$0.10	$0.50	Cheapest OpenAI tier, index 38

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Is open weight still a Chinese edge now that Meta is back?

Yes. American labs have returned to open weights, but their models still trail the best Chinese releases.

A company can download an open-weight model, run it where it wants and fine-tune it on private data. It can switch hosting providers without changing the model. Once AI sits deep inside a product, that control protects the customer from changes in price, behavior and provider policy.

The American answer arrived in October 2026. Meta released Muse Spark Open on October 15: 816 billion parameters, 42 billion active and a one-million-token context. Meta calls it the most capable US open-weight model to date. Reflection AI released Beam under Apache 2.0 in the same month.

Neither has caught up. Artificial Analysis scores Muse Spark Open at 39 and Beam at 30, against 44 to 46 for the leading Chinese open models. Meta’s own benchmarks show Muse Spark Open ahead of Kimi K3 and GLM-5.3 on some agent tests, but behind them on coding. On Terminal-Bench 2.1, Kimi K3 scores 85.0, GLM-5.3 83.9 and Muse Spark Open 76.4. OpenAI’s gpt-oss-120b scores 12.

Licenses still need checking model by model. DeepSeek and Z.ai have used MIT, and Qwen’s major open releases generally use Apache 2.0. Kimi K3 uses a custom license: large “model as a service” providers earning over $20 million a year need a separate deal with Moonshot. Meta’s Muse license has a similar $20 million clause and requires a “Built with Muse Spark Open” label.

The open-weight market is now a real contest, with an American side for buyers who want to avoid Chinese models. On quality, the Chinese side still leads it.

Can ordinary companies really run Chinese open models themselves?

Most companies can run smaller Chinese models, but the leading releases still require data-center hardware.

Kimi K3’s weights take about 1.5 terabytes, for 2.8 trillion parameters. DeepSeek V4.1 Flash has 552 billion parameters. Loading the weights is only the first step; serving many users quickly takes extra memory, several accelerators and specialized inference software.

That puts the top models beyond an ordinary company server. They are open in the sense that customers can control them, but most will still rent the hardware from AWS, Azure, Google Cloud or a specialist host.

Smaller models spread further. Qwen3.6-35B-A3B has 35 billion parameters and activates only three billion per token. Z.ai’s GLM-5.3 is reported to use 320 billion parameters with 18 billion active, smaller than GLM-5.2. A model of that kind can power coding, extraction, support and internal search without trillion-parameter costs.

American open models face the same limits. Meta says Muse Spark Open can run on two linked Mac Studios with 512 GB of memory each, a costly setup. OpenAI’s gpt-oss-20b runs within 16 GB, but it is far weaker.

The largest models get the headlines. The efficient mid-sized ones are more likely to end up inside thousands of companies.

Model	Approximate footprint or architecture	Who can realistically run it?
Kimi K3	2.8T parameters, about 1.5 TB of weights	Large data centers and cloud providers
Muse Spark Open	816B total, 42B active	Multi-GPU servers; two high-end Mac Studios at minimum
DeepSeek V4.1 Flash	552B total, 8B–16B active	Multi-GPU enterprise deployments
GLM-5.3	320B total, 18B active (reported)	Multi-GPU enterprise deployments
Qwen3.6-35B-A3B	35B total, 3B active	Smaller professional deployments
OpenAI gpt-oss-20b	Fits in about 16 GB	Workstations and some edge systems

Do developers now send more traffic to Chinese models than OpenAI?

On the largest open marketplace, yes. Across OpenAI’s whole platform, almost certainly not.

OpenRouter’s data for the week beginning October 12, 2026 is the clearest signal. DeepSeek had 24.4% of text requests, Google 19.7% and OpenAI 18.5%. Z.ai took 12.2% and Qwen 5.3%. Those three Chinese labs together handled 41.9% of requests, more than twice OpenAI’s share.

Token volumes point the same way. DeepSeek V4.1 Flash processed 29.6 trillion tokens on OpenRouter in the week to October 19, 2026. Seven of the ten most-used models were Chinese, including models from Xiaomi, StepFun and Tencent. GPT-6 Luna ranked eighth.

These figures need care. OpenRouter attracts cost-sensitive developers, and most OpenAI traffic goes through OpenAI’s own API, which OpenRouter does not see. A share of requests is not a share of revenue.

Downloads tell a longer story. Alibaba told Bloomberg in August 2026 that Qwen models had passed three billion downloads and more than 300,000 derivative models. That is a company figure, and one download is not one deployment, but derivatives show developers adapting the models, not just testing them.

Developers are also hedging. In a September 2026 Stacklok survey of 520 North American enterprises, about 92% used more than one model provider. Chinese models win a share of the workload rather than replacing OpenAI outright.

Have American cloud platforms already made Chinese models mainstream?

Yes. All three large American clouds now sell recent Chinese models as ordinary enterprise options.

AWS made Kimi K3 generally available on Amazon Bedrock on September 18, 2026, inside AWS security and governance controls. Microsoft Foundry lists DeepSeek V4.1 Flash, Kimi K3, GLM-5.3 and GLM-5.3 Flash, served through Fireworks. Google’s Vertex AI lists DeepSeek V4.1 Flash and GLM 5.2 among its managed open models.

The details differ by platform. Microsoft notes that Fireworks models on Foundry share data with the Fireworks service, so a company’s data terms depend on how the model is deployed. Legal teams need to check each route.

That distribution changes the trust question. A European or American company can use Chinese weights while keeping data, contracts and monitoring with a cloud provider it already trusts. The model comes from China; the controls do not.

Cloud support also lowers switching costs. Customers can compare models on one platform and move routine work to cheaper options without committing the company to DeepSeek or Qwen.

The clouds could have treated Chinese models as politically untouchable. Instead, they list them within weeks of release. For buyers, that is the strongest sign Chinese models have become normal infrastructure.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Does censorship make Chinese open models too risky?

Censorship makes Chinese open models unsuitable for some information-heavy uses, but it will not stop their adoption in coding and ordinary business automation.

Research has found systematic problems around politically sensitive subjects. One study of open-weight models from Qwen, DeepSeek and MiniMax found frequent refusals, deflections and false claims about the Tiananmen protests, Falun Gong and the treatment of Uyghurs. Models sometimes held the relevant knowledge but suppressed it in the final answer.

Another audit tested DeepSeek on 646 politically sensitive prompts. References to government accountability, transparency and civic mobilization were sometimes removed or softened between the model’s reasoning and its answer.

That is a serious problem for research tools, education, journalism and general assistants. A user may get an incomplete answer without knowing why.

The risk is much smaller for narrow tasks. Political censorship rarely affects a model that extracts invoice fields, fixes code or sorts support tickets. Companies can also host the weights, test them and fine-tune behavior.

Research reported by Semafor in July 2026 found that censorship in Chinese models can be undone, including in models distilled from them. Earlier work found that prompting and fine-tuning reduce censorship without removing every false answer. Open access makes the problem easier to fix, not automatically gone.

Chinese models will lose some sensitive Western customers over this. They can still win the large market where political questions never come up.

Will export controls on both sides slow Chinese models?

US chip controls slow Chinese labs but have not stopped them. China’s own proposed controls are a newer risk for foreign users.

US policy since January 2026 allows H200 exports to approved Chinese customers through case-by-case licenses, with 25% of sales going to the US government. Little has shipped. Nvidia said in August 2026 that its H200 sales to China in the quarter to late July were under 1% of its data-center revenue. Beijing had also discouraged purchases earlier in the year to support domestic chips, according to Reuters.

Chinese labs have adapted. Sparse activation, lower-precision computing and efficient attention get more work from each chip. DeepSeek V4.1 Flash activates only 8 billion of its 552 billion parameters on input. The models kept improving through 2026.

Compute still caps how far they can go. Meta is expanding Prometheus toward one million GPUs, and serving more than a billion weekly ChatGPT users takes far more capacity than publishing a strong checkpoint.

The new question is whether China will keep sharing. On October 16, 2026, its commerce and science ministries proposed export controls on 40 technologies, 23 of them in AI. According to published summaries, model weights and source code above certain performance thresholds would come into scope for the first time. Access through APIs and apps, and models already released as open source, would be exempt. Public comments run until November 15, 2026.

Reports differ on the threshold: one account cites models with two trillion or more parameters, while others say the thresholds are not quantified. Either way, the draft does not touch existing downloads. It could restrict future frontier releases, which is the scenario foreign buyers should plan for.

China can keep producing competitive models under US controls. Whether the best future ones will still be published openly is now a policy decision in Beijing.

Will Chinese models dominate inside China?

Chinese models will dominate mainland China because OpenAI cannot compete there through the same channels.

Mainland China remains absent from OpenAI’s official list of supported ChatGPT and API markets. OpenAI warns that accessing its products from unsupported locations can lead to an account being blocked or suspended.

Chinese providers operate inside a large domestic approval system. The Cyberspace Administration of China said 988 generative AI services had completed national filings by the end of June 2026. On September 14, 2026, it added 124 more services registered in July and August.

Domestic labs also bring Chinese-language performance, local cloud ties and integrations with large platforms. Alibaba can connect Qwen with its cloud and consumer services. MiniMax says its products reach more than 300 million users and 2 million enterprises and developers in total.

OpenAI may keep building stronger global models. That has limited value in a market where its access remains restricted.

China is large enough to support its own AI ecosystem. Chinese models can secure that market without first defeating OpenAI anywhere else.

Can Chinese open models break ChatGPT’s distribution moat?

No. Chinese open models cannot currently match ChatGPT’s consumer reach, and this is the biggest obstacle to beating OpenAI overall.

ChatGPT had more than 1.2 billion weekly active users on October 7, 2026, up from more than 900 million in March. OpenAI also reported more than 50 million paying consumer subscribers and more than nine million paying business seats in March 2026, its latest disclosed counts.

Those users keep conversations, files, routines and workplace connections inside ChatGPT. They do not leave because another model costs less per token.

OpenAI is turning that reach into new revenue. Codex passed five million weekly users in June 2026, and ChatGPT Ads reached a $1 billion annualized run-rate by August 31, 2026.

The moat is not untouchable, but the threat is American. Similarweb estimates ChatGPT’s share of visits to major AI chatbot websites fell from 78% to 57% in the year to August 2026. Claude was the main gainer, not a Chinese app.

Chinese adoption is fragmented. A Qwen or DeepSeek model may run inside a cloud, a private system or an app whose users never see its name. That spreads the technology without building a global customer relationship.

Open models can take work from OpenAI’s API. Taking users from ChatGPT is a different and much harder job.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Is OpenAI becoming vulnerable as models get commoditized?

Yes. OpenAI’s model advantage is weaker than ever, so its products must carry more of the business.

A customer once had to accept a large quality drop to leave the leading closed models. In October 2026, it can route requests among OpenAI, Anthropic, Google, Meta, DeepSeek, Z.ai, Kimi, Qwen and MiniMax by cost and quality.

Customers already do. Andreessen Horowitz found in January 2026 that 81% of surveyed organizations used three or more model families. Switching is not free: in Stacklok’s September 2026 survey, 57% estimated a move to another provider would take four months or more.

OpenAI’s September price cuts show the pressure. GPT-6.1 Sol and Luna cost about half their GPT-5.6 equivalents. Cheaper models protect market share but compress margins on every routine request.

The squeeze comes from two directions. Chinese open models set the price floor. Anthropic sets the pace in enterprise: it reportedly booked $11.6 billion in second-quarter revenue against OpenAI’s $6.7 billion, although the two count cloud-partner revenue differently.

The model is becoming one component of a larger system. Memory, integrations, security, user experience and distribution can keep a customer when a rival model is cheaper.

Chinese open models are removing the scarcity of advanced intelligence. OpenAI’s future depends on a product layer that is harder to copy than the model.

Can Chinese laboratories build businesses as powerful as OpenAI?

Not soon. Chinese labs are building real businesses, but the largest is roughly one-fiftieth of OpenAI’s size by revenue.

OpenAI reached annualized revenue of about $50 billion in September 2026, according to Reuters, and has pitched investors a $70 billion target for year-end. Its last completed round, in March 2026, valued it at $852 billion. In October 2026, it was reported to be seeking about $30 billion more at a valuation above $1.4 trillion; that round has not closed.

The Chinese labs are growing fast from a small base. DeepSeek’s annualized revenue passed $1 billion in September 2026, according to The Information, more than double the level of July. Z.ai reported RMB953.89 million in first-half revenue, with substantial losses. MiniMax reported $116.6 million, with an adjusted net loss of $293 million.

Investors are paying up. Moonshot AI completed a round at about $50 billion in October 2026, according to Bloomberg, and is preparing a possible Hong Kong IPO in early 2027. DeepSeek’s fundraising reportedly drew commitments above RMB80 billion against a RMB50 billion target, though The Information described a $7.5 billion raise; neither has been confirmed closed. Z.ai’s market value is estimated at about HK$227 billion.

The open model is also the business problem. DeepSeek leads OpenRouter’s traffic, but much of that traffic is served by other hosts, who keep the revenue. Qwen may power thousands of products while AWS or the app developer collects most of the money.

A Chinese lab could still become as influential as OpenAI. It would need a product or cloud platform that captures the value its models create.

What would Chinese open models need to win outright?

Chinese open models would need lasting technical leadership and a global product layer before we could say they had beaten OpenAI outright.

First, the frontier. A Chinese open family would need to reach the top of the independent rankings and stay there across coding, reasoning, agents and multimodal work. In October 2026, the best sits about 12 points below the leader on Artificial Analysis.

Second, infrastructure. Releasing weights lets others help with hosting, but direct enterprise relationships need a dependable API and support at global scale.

Third, trust. Western buyers need clear licenses, security documentation, predictable updates and honest answers on censorship. China’s draft export controls add a new question: whether future frontier weights will still leave the country.

Fourth, the interface. Models sold through Amazon, Microsoft or third-party apps may weaken OpenAI while enriching the cloud or the app developer.

Finally, the winner needs a full system for coding, research, memory, company data and real actions in software. OpenAI, Anthropic and Google already compete at that level.

Chinese labs meet much of the technical and pricing requirement. Their distribution, trust and global product position remain much weaker.

Will Chinese open models beat OpenAI?

Chinese open models are already beating OpenAI on capability per dollar and on open-marketplace traffic, but they are unlikely to beat OpenAI as a company anytime soon.

The model-layer case is strong. The five best open-weight models are Chinese. DeepSeek leads OpenRouter’s request share, ahead of OpenAI. AWS, Microsoft and Google all host recent Chinese models.

OpenAI has started to respond. GPT-6 Luna matches Chinese list prices, but with a weaker model. To match Chinese capability, OpenAI still charges many times more.

The frontier picture has changed too. In October 2026, Anthropic, Meta and Google sit at or above OpenAI on the main leaderboards, with Chinese open models about eight points below GPT-6 Astra. OpenAI is squeezed from both sides.

OpenAI keeps the advantage that is hardest to copy: more than 1.2 billion weekly ChatGPT users and around $50 billion in annualized revenue, against about $1 billion for DeepSeek. Chinese open models can replace many GPT calls without replacing ChatGPT.

Two risks could change the trajectory. China’s draft export controls could limit future open releases, and American open models from Meta and Reflection give Western buyers an alternative if they improve.

So the direct answer is partly yes. Chinese open models are compressing OpenAI’s prices, removing its scarcity and winning developer traffic. They are not on course to replace OpenAI’s global business.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.