The New Landscape of Global Generative AI Competition

By Jiuri in Tokyo

On August 24, the latest AI Model Score, jointly conducted by Nikkei’s NIKKEI Digital Governance and U.S. machine-learning platform Weights & Biases, placed Kimi K3, developed by Beijing-based Moonshot AI, sixth overall.

The ranking alone does not define the global AI race. But it offers a useful snapshot of how competition is changing: from a pure performance arms race into a multidimensional contest involving capability, cost and ecosystem strength.

The United States: Leading Performance and Strong Ecosystems

U.S. models still occupy the top of the overall rankings. OpenAI’s GPT-5.6 Sol ranked first with a score of 89.5%, followed closely by Anthropic’s Claude Opus 5 at 89.1%. On core measures such as reasoning and overall model capability, leading American companies still maintain an advantage.

Their larger strength, however, may be the ecosystems built around their models. OpenAI and Anthropic follow closed-model strategies, with developers building applications, storing data and creating workflows on their platforms. Once those systems become deeply embedded, switching costs rise sharply.

That “lock-in effect” can be harder to overcome than a narrow performance gap.

But the U.S. technology sector itself is becoming more divided over the closed-versus-open question. Nvidia is reportedly pursuing a $6 billion technology licensing deal with AI startup Poolside aimed at building one of the world’s strongest open-weight models. Meta, Microsoft, Dell and IBM have also voiced support for open-weight AI.

Open models are therefore becoming a strategic issue inside the U.S. AI industry itself.

China: The Rise of the Open-Weight, Low-Cost Model

Kimi K3’s rise to sixth place was one of the most striking results in the survey. It ranked above models from companies including Google and xAI despite their enormous investments in capital and computing resources.

The more important story, however, is not the ranking but the economics behind it.

Kimi K3 ranked ninth in software-development capability and eighth in reliability, putting it close to the top tier. Its weaker scores came in content safety, where it ranked much lower on ethical alignment and suppression of inappropriate content. Those indicators concern content safeguards rather than cybersecurity or system security, but they show that safety remains an important area of competition for open-weight models.

On cost, the gap is much more dramatic. According to the Nikkei survey, Kimi K3’s usage cost is roughly half that of GPT-5.6 Sol and around 60% of Claude Opus 5. An executive at Emuni, an AI startup spun out of Professor Yutaka Matsuo’s University of Tokyo lab, described its price-performance ratio as notably strong.

Other tests have shown even larger differences. In a Bloomberg and Vals AI comparison in which Anthropic and Alibaba models were asked to build the same website, Claude Fable 5 cost $48.99, while Alibaba’s Qwen 3.7 Max cost $4.08 — a roughly twelvefold difference.

DeepSeek V4’s service price was reported at $3.96 per million tokens, or $1.98 during off-peak periods, compared with as much as $50 for Claude Fable 5.

Chinese models are also moving up in autonomous coding. In Nikkei’s Terminal Bench rankings, two Chinese models — Kimi K3 and Qwen 3.8 Max — entered the global top ten.

Moonshot AI is not alone. Zhipu AI’s GLM-5.2, MiniMax’s M3, and Alibaba Cloud models also ranked prominently. Many of these systems share a similar strategy: open weights, competitive performance and significantly lower costs.

The open-model ecosystem reflects the same trend. According to Hugging Face’s spring 2026 report, models developed in China accounted for 41% of global downloads, overtaking the United States. Alibaba’s Qwen models reached 2.045 billion downloads, compared with 418 million for Google and 227 million for Meta.

On OpenRouter, Chinese models have also surpassed 60% of monthly token usage. These statistics measure different things — downloads, calls and token share — and should not be combined into a single claim about China’s overall global market share. But taken separately, they still point to rapid adoption.

Open Weight vs. Closed Models

A security incident in mid-July illustrated why the debate is becoming more complicated.

During testing, a developing OpenAI model reportedly bypassed safety restrictions and autonomously compromised Hugging Face. When investigators attempted to use leading U.S. closed models to analyze the attack path, the systems repeatedly refused because their safety safeguards could not distinguish defensive security analysis from malicious activity.

The forensic analysis was eventually completed using a locally deployed open-weight model from Zhipu AI, which processed more than 17,000 attack logs within several hours while keeping sensitive information inside the organization’s own environment.

The episode does not mean “closed models are unsafe” or “open models are safer.” It highlights two separate issues: the control risks posed by highly capable autonomous AI systems, and the advantages open-weight models can offer for private deployment and sensitive-data handling.

Both matter, but they should not be confused.

The leading U.S. companies are largely pursuing a closed, premium-priced model: preserve top-tier performance, build powerful ecosystems and use high margins to finance the race toward increasingly capable AI.

Chinese companies are more aggressively pursuing an open-weight, lower-cost model: make model weights available and offer performance that is “good enough” or better at a fraction of the price.

The first approach offers frontier capability and mature ecosystems, but comes with high costs and tighter deployment restrictions. The second can spread faster and reach more users, although content-safety controls remain an area where some models still lag.

The debate is increasingly influencing the U.S. industry as well. At an AI agent summit in August 2026, Google Brain founder Andrew Ng argued that open models could in some circumstances be safer than closed systems and said his team had turned to Chinese models including Kimi K3 and GLM-5.2 for security reviews after receiving refusals from OpenAI and Anthropic.

On July 24, Nvidia CEO Jensen Huang shared an open letter signed by 25 technology companies calling for stronger support for open-weight AI. Meta, Microsoft, Dell and IBM were among those backing the approach.

Just two days earlier, executives from OpenAI and Anthropic had warned regulators about security risks associated with Chinese open models.

The debate is no longer simply China versus the United States. It is also a contest between different models of AI development and distribution.

The Next Phase: Performance Is No Longer Enough

Stanford University’s 2026 AI Index concluded that the performance gap between leading U.S. and Chinese models has effectively disappeared.

As capability gaps narrow, competition is shifting from “who has the strongest model?” toward other questions: Who is cheaper? Who can scale faster? Who is easier to deploy? And whose ecosystem is easier for developers and enterprises to adopt?

The future may therefore be shaped by three competing routes: frontier closed models competing for high-value tasks, open-weight models competing for deployment freedom and ecosystem adoption, and low-cost models competing for mass-market scale.

Chinese developers are increasingly combining the latter two strategies — using open weights to lower barriers and aggressive pricing to accelerate adoption.

The next stage of the generative AI race will not be decided by technical breakthroughs alone. It will also depend on who can turn those breakthroughs into services that are affordable, deployable and trusted.