Will Chinese open models beat OpenAI?
In our generative AI market deck, you will find everything you need to understand the market
SUMMARY
Will Chinese open models beat OpenAI? At the model layer, probably yes; as a global company and consumer platform, not anytime soon.
The threat is now structural rather than tied to one DeepSeek moment. Alibaba, Z.ai, Moonshot AI, MiniMax and DeepSeek are improving across different model generations and workloads, which makes a one-off fluke much harder to argue.
OpenAI still owns the strongest top model in the comparison, but the gap has narrowed enough to change buying behavior. Many companies do not need the absolute best score; they need a model that clears their task reliably at a reasonable cost.
Price may matter more than a few benchmark points. DeepSeek’s published API prices sit far below GPT-5.6 Sol, creating room for companies to reserve OpenAI for difficult work and route routine requests to cheaper Chinese models.
Open weights also change the balance of power. Customers can host, modify and move a model across infrastructure providers, which gives them more control than a closed API and weakens the lock-in that once protected frontier-model vendors.
The largest Chinese checkpoints still demand data-center hardware, so “open” does not automatically mean easy to run. The more disruptive releases may be smaller sparse models that activate only a few billion parameters and can fit into ordinary professional deployments.
Developer adoption is no longer theoretical. Qwen’s download and derivative-model counts, DeepSeek’s availability across hosting providers, and support from AWS, Microsoft and Google show that Chinese models are becoming normal infrastructure rather than exotic alternatives.
Censorship remains a real weakness for research, education, journalism and general assistants, while chip controls still limit how many frontier experiments Chinese laboratories can run. Neither constraint is likely to stop adoption in coding, extraction, support automation and other narrow business tasks.
China’s domestic market is already tilted toward local models because OpenAI lacks official distribution there. Globally, though, ChatGPT’s user base, workplace presence and product ecosystem remain much harder to copy than a benchmark lead or a cheap API.
The likely outcome is a divided market: Chinese open models take a growing share of the intelligence underneath applications, while OpenAI fights to own the products, agents and customer relationships above them. That would still be a major defeat for OpenAI’s old model moat, even if OpenAI remains the larger company.
This market map, featured in our generative AI market deck, highlights top companies and startups in the generative AI market
Why is this question serious now?
Chinese open models have become a serious threat because several laboratories are now producing competitive models, rather than one company delivering one surprising release.
DeepSeek first changed expectations by showing that a Chinese laboratory could build strong reasoning models under tighter hardware constraints. Since then, Alibaba’s Qwen team, Z.ai, Moonshot AI and MiniMax have kept the pressure on across coding, multimodal work, long contexts and agents.
The latest independent rankings show how crowded the frontier has become. Artificial Analysis currently counts six laboratories with models scoring above 50 on its Intelligence Index, compared with only two in early June. Z.ai’s GLM-5.2 leads the open-weight category with a score of 51, while MiniMax M3 and DeepSeek V4 Pro score 44.
Moonshot’s Kimi K3 sits even closer to the absolute frontier, scoring 57 against 59 for OpenAI’s strongest GPT-5.6 Sol setting. Kimi K3 is currently classified as proprietary by Artificial Analysis, so it should not be counted as an open-model victory. It still shows how far Chinese laboratories can now push capability. Moonshot’s Kimi K2.5 and K2.6 provide the open-weight part of that story.
The pattern now stretches across several companies and model generations. A temporary fluke is getting much harder to argue.
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 offering similar or better intelligence on the tasks users care about. That lead could disappear when OpenAI releases another model, so one leaderboard result should not be treated as a permanent win.
An ecosystem victory would be more durable. Chinese models would become the models developers download, modify, host and place inside their own products. Companies would build tools, training pipelines and internal expertise around them.
Beating OpenAI overall requires much more. OpenAI currently operates a consumer product with more than 900 million weekly users and over 50 million paying subscribers. It also sells APIs, workplace products, coding tools and agent systems to more than one million business customers.
For this article, we will call it a model-layer victory when Chinese open models combine near-frontier quality with better prices and greater customer control. An outright victory would require those advantages to produce stronger global distribution, products and revenue than OpenAI.
The first outcome is already plausible. The second remains distant.
If you want more recent data on this point, please see our latest generative AI market report.
As this chart shows, and as featured in our generative AI market deck, search interest in LLMs has surged
Are Chinese open models as smart as OpenAI now?
The best Chinese open model is currently close enough to OpenAI for many buyers, although OpenAI still has the stronger top model.
Artificial Analysis gives GPT-5.6 Sol a score of 59 on its current Intelligence Index. GLM-5.2, the highest-ranked open-weight model, scores 51. That eight-point gap is real, especially on the hardest coding, research and agent tasks. It is far smaller than the gap companies faced when open models were obviously second-tier.
GLM-5.2 also shows that the distance changes by task. Z.ai reports strong results in long-horizon coding and tool use, including 62.1% on SWE-bench Pro and 81% on Terminal Bench 2.1 under its evaluation setup. Vendor benchmarks require caution, but the model is clearly aimed at real software work rather than simple chatbot questions.
DeepSeek V4 Pro and MiniMax M3 sit further behind on the overall index. Yet both can still handle a large share of routine coding, analysis, extraction and agent work. Most businesses do not buy the highest possible benchmark score. They buy a system that performs their task reliably without making the bill or infrastructure unreasonable.
Kimi K3 sharpens the point, even though it is currently proprietary. It trails GPT-5.6 Sol by only two points overall. However, Artificial Analysis also found that its hallucination rate rose from 39% for Kimi K2.6 to 51% for K3. A near-equal headline score can still hide an important weakness.
Chinese open models have therefore reached the range where testing them makes commercial sense. They have not reached the point where we can simply call them better.
| Model | Availability | Intelligence Index | Gap from GPT-5.6 Sol |
|---|---|---|---|
| GPT-5.6 Sol | Closed API | 59 | — |
| Kimi K3 | Currently proprietary | 57 | 2 |
| GLM-5.2 | Open weight | 51 | 8 |
| DeepSeek V4 Pro | Open weight | 44 | 15 |
| MiniMax M3 | Open weight | 44 | 15 |
| OpenAI gpt-oss-120b | Open weight | 24 | 35 |
Are Chinese laboratories closing the gap faster than OpenAI can reopen it?
Chinese laboratories are closing the gap quickly, but OpenAI can still reopen it whenever a strong new generation arrives.
The current race looks like a repeated cycle. OpenAI or another American laboratory pushes the frontier forward. Chinese models then approach that level more quickly and cheaply than before. The distance shrinks until the next frontier release.
The pace has accelerated lately. Artificial Analysis recorded four major frontier releases within eight days, including GPT-5.6 and Kimi K3. Six laboratories now have a model scoring above 50 on its index. Frontier capability is spreading across more teams instead of staying concentrated inside one or two companies.
Chinese development is also broadening. DeepSeek V4 supports million-token contexts. GLM-5.2 targets long-running coding and agent work. Moonshot’s Kimi K2.6 combines multimodal inputs with long-horizon execution, while its newer K2.7 Code model cuts thinking-token use by around 30% compared with K2.6.
OpenAI still holds two advantages. It can run many expensive experiments at once, and it can immediately place a successful model in front of hundreds of millions of users. A Chinese laboratory may match one model while OpenAI learns from usage at a much greater scale.
What has changed is how long the lead lasts. OpenAI can still move ahead, but Chinese competitors are spending less time far behind.
This chart, featured in our generative AI market deck, shows annual VC investment in generative AI startups
Are Chinese open models already much cheaper?
Chinese open models are dramatically cheaper than OpenAI’s frontier API today, even after allowing for differences in quality.
OpenAI charges $5 per million input tokens and $30 per million output tokens for GPT-5.6 Sol. GPT-5.6 Terra costs $2.50 and $15, while Luna costs $1 and $6.
DeepSeek charges $0.435 per million uncached input tokens and $0.87 per million output tokens for V4 Pro. V4 Flash costs $0.14 for uncached input and $0.28 for output through DeepSeek’s own API.
At those prices, GPT-5.6 Sol costs roughly 11 times more than V4 Pro for input and 34 times more for output. Against V4 Flash, the difference grows to around 36 times for input and 107 times for output.
Those ratios exaggerate the practical saving when the cheaper model needs more retries or produces weaker answers. Even so, the price difference is far too large to dismiss. A company running millions of routine requests can send harder work to OpenAI and route the rest to DeepSeek or another open model.
The cost advantage also reflects real engineering. DeepSeek V4 Pro contains around 1.6 trillion total parameters but activates roughly 49 billion for each token. V4 Flash activates about 13 billion out of 284 billion. Sparse architectures let the model hold a lot of knowledge while using only part of the network for each request.
Public pricing does not tell us whether every cheap endpoint earns a healthy margin. Providers may discount access to gain users or fill spare capacity. But the efficiency work is real, so subsidies cannot explain the whole gap.
| Model | Input per 1M tokens | Output per 1M tokens | Overall position |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | OpenAI’s strongest model |
| GPT-5.6 Terra | $2.50 | $15.00 | Mid-priced frontier option |
| GPT-5.6 Luna | $1.00 | $6.00 | Cheapest GPT-5.6 tier |
| DeepSeek V4 Pro | $0.435 | $0.87 | Strong Chinese open model |
| DeepSeek V4 Flash | $0.14 | $0.28 | High-volume, lower-cost option |
If you want more recent data on this point, please see our latest generative AI market report.
Does open weight actually give Chinese models an edge?
Open weights give Chinese models a clear edge when companies want control, customization or freedom from one API provider.
A company can download an open-weight model, run it in its chosen environment and fine-tune it on private data. It can inspect the deployment, change the surrounding safety rules and switch infrastructure providers without replacing the model itself.
That flexibility becomes valuable once AI sits deep inside a product. A company that relies entirely on a closed API must accept changes in price, availability, model behavior and provider policy. Open weights give the customer more leverage.
Several important Chinese releases also use permissive licenses. GLM-5.2 uses MIT and explicitly states that it has no regional restrictions. DeepSeek V4 Pro is also released under MIT. Qwen’s major open releases generally use Apache 2.0, including Qwen3.6-35B-A3B.
The word “open” still needs checking model by model. Some releases offer weights without training data or complete training code. Others add commercial restrictions. A legal team should read the actual license rather than assuming every downloadable Chinese model grants the same rights.
OpenAI has already responded. Its gpt-oss-120b and 20b models use Apache 2.0 and can be customized or hosted independently. The larger model fits within 80 GB of memory, while the smaller one requires about 16 GB.
OpenAI’s move confirms the pressure Chinese models have created. The company would have little reason to return to open-weight language models if that part of the market were irrelevant.
This chart, featured in our generative AI market deck, looks at OpenAI’s strategy in generative AI
Can ordinary companies really run Chinese open models themselves?
Most companies can run smaller Chinese models, but the largest frontier releases still require serious data-center hardware.
GLM-5.2’s published files total about 1.51 terabytes. DeepSeek V4 Pro’s model files are roughly 865 gigabytes. Loading the weights is only the first step; serving many users quickly requires additional memory, several accelerators and specialized inference software.
That puts the biggest models beyond an ordinary company server. They are open in the sense that customers can control them, although many customers will still rent the hardware from AWS, Azure or another provider.
Smaller models are far more disruptive. Qwen3.6-35B-A3B has 35 billion total parameters while activating only three billion for each token. Alibaba offers it as open weights and through an API. A model of that size can power coding, extraction, customer support and internal search without the cost of a trillion-parameter system.
Moonshot has followed a similar path. Kimi Linear uses 48 billion total parameters with around three billion active parameters. Moonshot reports up to 75% lower key-value cache use and up to six times faster decoding in long-context tests. Those figures come from its own research, but they show where the engineering effort is going.
OpenAI has also made compact deployment easier through gpt-oss. Its 20-billion-parameter model activates 3.6 billion parameters per token and can run within 16 GB of memory.
The largest models get the headlines. Smaller models may spread further because thousands of companies can actually afford to use them.
| Model | Approximate footprint or architecture | Who can realistically run it? |
|---|---|---|
| GLM-5.2 | About 1.51 TB of files | Large data centers and cloud providers |
| DeepSeek V4 Pro | About 865 GB of files | Multi-GPU enterprise deployments |
| Qwen3.6-35B-A3B | 35B total, 3B active | Smaller professional deployments |
| OpenAI gpt-oss-120b | Fits in about 80 GB | One high-end enterprise GPU |
| OpenAI gpt-oss-20b | Fits in about 16 GB | Workstations and some edge systems |
Are developers actually choosing Chinese open models?
Developers are already adopting Chinese open models at huge scale, although OpenAI still handles much more usage across its complete platform.
Qwen offers the clearest evidence of ecosystem reach. Alibaba says the family passed one billion cumulative downloads on Hugging Face and has produced more than 200,000 derivative models. The company has released more than 400 Qwen models since 2023.
Those figures should be read carefully. One download does not equal one developer or one production deployment. Automated systems may download the same model repeatedly, and derivative models range from serious commercial adaptations to small experiments.
The scale is still hard to fake. More than 200,000 derivatives means developers are adapting Qwen for languages, industries, devices and specific tasks rather than merely testing the original checkpoint once.
DeepSeek shows demand through API marketplaces. OpenRouter lists V4 Pro with 17 hosting providers and V4 Flash with 19, giving developers many ways to use the same weights without operating them directly.
OpenAI remains much larger as a platform. Its APIs currently process more than 15 billion tokens each minute. Maintained for 30 days, that pace would equal roughly 648 trillion tokens. That figure covers the whole OpenAI platform, so comparing it with one Chinese model or one marketplace would be misleading. It does show the size of OpenAI’s existing developer engine.
Chinese models have passed the adoption test. They have not yet passed OpenAI in total usage.
This chart, featured in our generative AI market deck, shows annual funding in generative AI startups
Have American cloud platforms already made Chinese models mainstream?
American cloud companies have already turned Chinese open models into normal options for Western developers.
Amazon Bedrock added fully managed versions of DeepSeek V3.2, MiniMax M2.1, GLM 4.7, Kimi K2.5 and Qwen3 Coder Next. Customers can use them through Amazon’s infrastructure without building their own inference clusters.
Microsoft includes DeepSeek in Azure Foundry alongside OpenAI, Meta, Mistral and other model providers. Google offers DeepSeek and Qwen models through its cloud model catalog.
That distribution changes the trust equation. A European or American company can use Chinese weights while keeping its data and infrastructure inside an American cloud environment. The model comes from China, while the hosting, contracts, access controls and monitoring come from a provider the company already uses.
Cloud support also removes much of the operational pain. Customers can compare models through one platform and move routine workloads to cheaper options. They do not need to commit the whole company to DeepSeek or Qwen.
The large American clouds could have treated Chinese models as niche or politically untouchable. Instead, they are selling them. Developer demand has become too large to ignore.
If you want more recent data on this point, please see our latest generative AI market 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.
Recent research has found systematic problems around politically sensitive subjects. One study examined open-weight models from Qwen, DeepSeek and MiniMax and found frequent refusals, deflections and false claims about topics including the Tiananmen protests, Falun Gong and the treatment of Uyghurs. The researchers also found that models sometimes contained the relevant knowledge while suppressing it in the final response.
Another audit tested DeepSeek on 646 politically sensitive prompts. It found that references to government accountability, transparency and civic mobilization were sometimes removed or softened between the model’s reasoning and final answer.
That creates a serious problem for research tools, education, journalism and general assistants. A user may receive an incomplete answer without knowing that the omission came from political training rather than a lack of knowledge.
The risk changes for narrower applications. Political censorship will rarely affect a model that extracts invoice fields, fixes code or sorts customer-service tickets. Companies can also host the weights themselves, test the model and fine-tune some behavior.
Researchers have managed to reduce censorship through prompting and fine-tuning, but no tested method eliminated false responses completely. Open access makes the problem easier to study and modify without making it disappear.
Chinese models will lose some sensitive Western customers because of this weakness. They can still win a large commercial market where political questions barely appear.
This chart, featured in our generative AI market deck, compares the main business model options for generative AI SaaS platforms
Will chip controls stop China from keeping up?
Chip controls will slow Chinese model development, but they have already failed to stop Chinese laboratories from reaching the frontier.
The United States continues to restrict access to advanced computing hardware and semiconductor technology. Its current policy allows some H200, AMD MI325X and similar exports to approved Chinese customers through case-by-case licenses, rather than providing unrestricted access.
That leaves Chinese laboratories with less predictable access to the best training and inference hardware. The constraint becomes especially painful when a new model attracts more users than its creator can serve.
Yet the model releases have continued. DeepSeek, Qwen, GLM, Kimi and MiniMax have all improved while those restrictions were in force. Sparse activation, lower-precision computation and more efficient attention let laboratories obtain more work from each chip.
GLM-5.2, for example, uses a shared sparse-attention index that Z.ai says reduces per-token computation by 2.9 times at a one-million-token context. DeepSeek’s newer architectures also focus heavily on lowering the memory and compute cost of long sequences.
Export controls still give OpenAI an important advantage at extreme scale. Training many frontier experiments and serving nearly a billion weekly users require far more compute than publishing a strong checkpoint.
China can keep producing competitive models under these constraints. Matching OpenAI’s total capacity will be harder.
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 says 988 generative AI services had completed national filings by the end of June, alongside 598 registered applications or features using filed models. Another 120 services were added during the previous two-month reporting period alone.
Domestic laboratories also bring Chinese-language performance, local cloud relationships and integrations with the country’s largest technology platforms. Alibaba can connect Qwen with its cloud and consumer services. Other laboratories can work with domestic phone makers, software companies and state-linked customers.
OpenAI may continue to produce a stronger global model. That advantage has limited value in a market where its official access and distribution remain restricted.
China is large enough to support an independent AI ecosystem. Chinese models can secure that market without first defeating OpenAI everywhere else.
This chart, featured in our generative AI market deck, illustrates how revenue is distributed across customer segments in the generative AI market
Can Chinese open models break ChatGPT’s distribution moat?
Chinese open models cannot currently match ChatGPT’s consumer reach, and this is the biggest obstacle to beating OpenAI overall.
ChatGPT has more than 900 million weekly active users and over 50 million paying consumer subscribers. OpenAI also reports more than seven million workplace seats and more than one million business customers.
Those users have saved conversations, files, routines and workplace connections inside ChatGPT. They do not automatically leave because another model gains a few benchmark points or costs less per token.
OpenAI is also turning its distribution into new products. Codex has reached three million weekly active users, while enterprise products now generate more than 40% of company revenue.
Chinese model adoption is more fragmented. A Qwen model may run inside an Alibaba service, an American cloud, a private enterprise system or a product whose users never see the Qwen name. That spreads the technology without creating one global customer relationship.
Alibaba has the best chance of changing this because it owns cloud, commerce and consumer platforms. Even so, no Chinese AI application currently approaches ChatGPT’s worldwide habit and brand recognition.
Open models can take work away from OpenAI’s API. Taking users away from ChatGPT will be much harder.
If you want more recent data on this point, please see our latest generative AI market report.
Is OpenAI becoming vulnerable as models get commoditized?
OpenAI’s model advantage is becoming less defensible, so the company now needs its products to carry more of the business.
A customer once had to accept a large quality drop to leave the leading closed models. Today it can choose among OpenAI, Anthropic, Google, DeepSeek, Qwen, GLM, Kimi and MiniMax, then route each request to the best mix of quality and cost.
This weakens premium pricing. A business may reserve GPT-5.6 for hard research or high-value coding while using a cheaper open model for extraction, classification and ordinary support work.
OpenAI’s recent strategy reflects that pressure. It brought open-weight language models back through gpt-oss, while continuing to expand ChatGPT, Codex, enterprise agents and connected workplace tools.
The model increasingly serves as one component inside a broader system. Memory, integrations, evaluation, security, user experience and distribution can keep a customer even when a rival model becomes cheaper.
Chinese open models are already reducing the scarcity of advanced intelligence. OpenAI’s future depends on making the surrounding product harder to copy than the model itself.
This chart, featured in our generative AI market deck, shows how AI video generation technology has evolved over time
Can Chinese laboratories build businesses as powerful as OpenAI?
Chinese laboratories can build major AI businesses, but open-model popularity alone will not produce OpenAI-sized revenue.
OpenAI currently generates around $2 billion in revenue each month and has raised approximately $122 billion. Its income comes from consumer subscriptions, enterprise seats, APIs, coding products and other services surrounding its models.
An open-model creator gives customers more ways to avoid paying it directly. A developer can download the weights, use a competing host or modify the model without maintaining a long-term relationship with the original laboratory.
That does not make open models bad businesses. Alibaba can treat Qwen as a way to sell more cloud infrastructure and protect itself from dependence on another AI provider. DeepSeek, Moonshot, Z.ai and MiniMax can charge for hosted access, premium models, enterprise support and specialized products.
The challenge lies in capturing enough of the value. Qwen may power thousands of successful companies while AWS, an application developer or the end customer earns most of the resulting revenue.
OpenAI keeps tighter control over the customer and the paid product. That structure currently makes it easier to turn technical progress into cash.
A Chinese company could still become as influential as OpenAI. It will probably need a powerful product or cloud platform alongside the open model.
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.
A temporary leaderboard lead would be insufficient. At least one Chinese open family would need to stay near the top for several generations, across coding, reasoning, agents and multimodal work.
The provider would also need enough infrastructure to serve global demand reliably. Releasing the weights lets other companies help with hosting, but the original laboratory still needs a dependable API and support system if it wants direct enterprise relationships.
Trust would have to improve as well. Western companies need clear licenses, security documentation, predictable updates and honest answers about political censorship and training behavior.
The hardest step would be owning the interface. Models distributed through Amazon, Microsoft or independent software products may weaken OpenAI while enriching the cloud company or application developer.
Finally, the winner would need a broader system for coding, research, memory, company data and real actions across software. OpenAI is already fighting on that level rather than relying only on model benchmarks.
Chinese laboratories currently meet much of the technical requirement. Their distribution, trust and global product position remain much weaker.
In our generative AI market deck, we identify pain points entrepreneurs should prioritize
Will Chinese open models beat OpenAI?
Chinese open models will probably beat OpenAI at the model layer, but they are unlikely to beat OpenAI as a company anytime soon.
The model case is strong. GLM-5.2 currently leads the open-weight rankings. DeepSeek offers capable models at a fraction of OpenAI’s API prices. Qwen has crossed one billion downloads and generated more than 200,000 derivatives. American cloud platforms now distribute Chinese models as ordinary enterprise options.
That combination will make Chinese models standard infrastructure for private deployments, customized business systems and cost-sensitive products. They will also dominate inside mainland China, where OpenAI lacks official distribution.
OpenAI’s technical moat is already weaker. It must compete against several Chinese laboratories, lower prices and support open weights while continuing to fund much larger training and serving operations.
OpenAI still owns the harder-to-copy advantage: a product used by more than 900 million people each week, millions of workplace users and a huge API ecosystem. Chinese open models can replace many GPT calls without replacing ChatGPT.
The likely result is a divided market. Chinese open models will take a large share of the intelligence running underneath applications. OpenAI will try to control the applications, agents and customer relationships above them.
So the direct answer is partly yes. Chinese open models are on course to weaken OpenAI’s model leadership, compress its prices and remove much of its technological scarcity. They are not currently on course to replace OpenAI’s global business.
If you want more recent data on this point, please see our latest generative AI market report.
OUR METHODOLOGY
This analysis tests whether Chinese open models are on course to beat OpenAI. Because that question mixes technical performance, developer adoption, economics, distribution and company strength, we broke it into separate analytical dimensions before bringing the evidence back together.
We examined model capability, pricing, licensing, deployment requirements, developer adoption, cloud distribution, censorship risk, chip constraints, China’s domestic market and OpenAI’s product and business moat. This separates a model-layer victory from an ecosystem victory and from the much harder claim that a Chinese company will overtake OpenAI overall.
For each dimension, we prioritized the freshest evidence available. We aggregated recent model releases, independent benchmark results, official API pricing, license terms, model files, cloud-platform catalogs, developer adoption data, company disclosures and regulatory information. One launch or one benchmark was never allowed to carry the conclusion by itself.
We used Artificial Analysis as the clearest cross-model capability benchmark because it puts open-weight and proprietary models on one current index. Vendor benchmarks from Z.ai, DeepSeek, Moonshot and others were used to understand task-level strengths, especially coding, tool use and long-context work, rather than as direct substitutes for independent evaluation.
Pricing comparisons use published API list prices for input and output tokens. They show the size of the current price gap, not the full cost of production deployment. Retries, latency, output quality, hosting, engineering work and utilization can change the practical economics.
We treated “open” model by model. Downloadable weights, permissive licenses and self-hosting rights were considered meaningful advantages, while the absence of training data or complete training code was not ignored. MIT and Apache 2.0 releases were checked separately rather than assuming every Chinese model offers the same rights.
Developer adoption was assessed through several forms of evidence: Hugging Face downloads, derivative-model counts, OpenRouter hosting availability and inclusion in the model catalogs of major American clouds. These indicators measure different things, so we used them together to distinguish broad experimentation from genuine distribution.
OpenAI’s business position was assessed separately from model quality. Consumer reach, paying subscribers, workplace seats, business customers, API usage, product adoption and revenue were used to judge whether technical pressure at the model layer is translating into a threat to the company as a whole.
The final conclusion is not produced by a weighted score. It reflects the balance of evidence across the dimensions above, with more confidence placed on official pricing, licenses, cloud listings and independent evaluations than on promotional claims or isolated benchmark wins.
Key sources used for this analysis include: Artificial Analysis and its Intelligence Index, OpenAI’s GPT-5.6 announcement, OpenAI API pricing, OpenAI model documentation, OpenAI’s supported countries and territories, DeepSeek API pricing, DeepSeek’s official GitHub organization, Qwen’s official GitHub organization, the Qwen blog, Z.ai’s official GitHub organization, Moonshot AI’s official GitHub organization, OpenRouter’s model directory, Amazon Bedrock’s supported-model catalog, Azure AI Foundry’s model catalog, Google Cloud’s Model Garden, Qwen on Hugging Face, DeepSeek on Hugging Face, the Cyberspace Administration of China, SWE-bench, and Terminal-Bench.
This chart, featured in our generative AI market deck, illustrates how revenue is distributed across Europe, Asia, North America, Africa, and South America in the generative AI market
