Chinese AI Models Close in on OpenAI as Google Delays Gemini 3.5 Pro

Google's Gemini 3.5 Pro has reportedly been delayed due to performance issues, while Chinese AI companies have unveiled a new generation of frontier-class models that are rapidly narrowing the gap with leading US systems.

Moonshot AI recently introduced Kimi K3, a 2.8-trillion-parameter model, while Alibaba unveiled a preview of Qwen 3.8 Max, its flagship 2.4-trillion-parameter multimodal model.

Kimi K3, in particular, has attracted attention from international media after posting benchmark results approaching those of Anthropic's Fable 5 and OpenAI's GPT-5.6. On several benchmarks, it outperformed GPT-5.6 Sol, OpenAI's highest-end reasoning model. Reuters quoted analysts as saying that the long-held view that Chinese AI models lag US frontier models by roughly six months is becoming increasingly difficult to defend.

Moonshot AI plans to release Kimi K3's model weights on July 27. In Artificial Analysis (AA)'s Intelligence Index, an independent benchmark of leading AI models, Kimi K3 ranked third with a score of 57, behind only Anthropic's Fable 5 and GPT-5.6, surpassing Gemini 3.0 Pro, Claude Opus 4.8, GPT-5.5 (xhigh), Claude Sonnet 5 and GLM-5.2.

The model demonstrated particularly strong performance in agentic AI and coding. It ranked fourth in AA's Coding Agent Index, trailing only GPT-5.6 Sol, GPT-5.6 Terra and Fable 5.

Kimi K3's overall ranking exceeded its coding-agent ranking largely because of its relatively weak performance on the SWE-Atlas-QnA benchmark.

AA's Coding Agent Index is calculated as the average of three benchmarks: Terminal-Bench v2, DeepSWE and SWE-Atlas-QnA. Kimi K3 scored 84% and 64% on the first two benchmarks, placing it among the top performers, but achieved only 23% on SWE-Atlas-QnA, significantly reducing its overall average.

Even so, Kimi K3 outperformed Fable 5 on Terminal-Bench v2 and exceeded SpaceXAI's Grok 4.5 and Claude Opus 4.8 on DeepSWE. It also ranked first on Frontend Code Arena, outperforming GPT-5.6.

The model also posted leading results on several agent benchmarks. It narrowly surpassed GPT-5.6 Sol on τ³-Banking, which evaluates tool use in multi-turn customer service scenarios modeled after retail banking environments.

Kimi K3 also ranked first on AutomationBench-AA, which evaluates SaaS workflow automation, and SpreadsheetBench 2, which measures spreadsheet task performance. On BrowseComp, a benchmark for web search and information retrieval, it achieved a state-of-the-art score of 91.2, exceeding GPT-5.6 Sol's previous record of 90.4.

The model ranked second on AA-Briefcase, CharXiv (RQ) with Tool, and Zerobench with Tool, and third on GDPval-AA v2, which measures workplace task performance across 44 occupations.

Its performance was comparatively weaker on several of the most demanding benchmarks. Kimi K3 ranked sixth on Humanity's Last Exam (HLE), behind Gemini Pro 3.1 Preview, and also placed sixth on CritPt, a benchmark evaluating reasoning over unpublished physics research. The top four positions on CritPt were occupied by GPT models, followed by Fable 5.

High performance comes with higher costs

Artificial Analysis noted that Kimi K3 remains relatively expensive to use because it is currently accessible only through Moonshot AI's API ahead of the release of its open weights. In addition, the model does not currently allow users to disable reasoning, meaning reasoning tokens are always generated and billed.

According to AA, Kimi K3 also tends to produce unusually verbose responses. Across the entire Intelligence Index evaluation, the model generated approximately 130 million output tokens, more than double the benchmark average of 63 million. The higher token usage substantially increases inference costs.

Once Kimi K3 becomes available as an open-weight model and third-party hosting providers begin offering it, market competition could reduce pricing. However, its enormous 2.8-trillion-parameter scale is likely to keep costs well above those of DeepSeek. Although Kimi K3 uses a Mixture-of-Experts (MoE) architecture that activates only 16 of 896 experts during inference, limiting computational requirements relative to its total parameter count, the GPU memory and compute resources required to serve the model remain substantial.

Alibaba introduces Qwen 3.8 Max

Alibaba unveiled a preview of Qwen 3.8 Max at the World Artificial Intelligence Conference (WAIC) in Shanghai on July 19.

The new model contains 2.4 trillion parameters, representing a significant increase from its predecessor. It is a large multimodal model capable of processing text, images, video and documents within a single system. Alibaba said it also plans to release the model weights as open source in the near future.

The company described Qwen 3.8 Max as its flagship model for coding and AI agents while claiming that its general capabilities rank just behind Fable 5. Alibaba highlighted significant improvements in agentic coding, software development, front-end development and function calling, as well as stronger tool use, complex workflow execution and multi-step task automation.

Alongside the preview release, Alibaba integrated Qwen 3.8 Max into its Token Plan, Qoder and QoderWork platforms.

Because the model was announced only recently, independent benchmarking organizations such as Artificial Analysis have not yet published evaluation results. However, early community testing on platforms including Aider Leaderboard, LiveBench and WebDev Arena suggests that the model is competitive with GPT-5.6 and Kimi K3 in front-end development, agentic coding and web user interface generation.

Weights & Biases (W&B) had previously ranked the earlier Qwen 3.7 Max among the top models for spreadsheet reasoning and workflow automation, raising expectations for the new release.

Qwen 3.8 Max also employs an MoE architecture, although Alibaba has not disclosed the number of active parameters. That figure will play a key role in determining inference costs. Qwen 3.7 Max was priced at $2.50 per million input tokens and $7.50 per million output tokens, but the substantially larger 3.8 Max model could carry higher pricing.

The release of increasingly competitive Chinese models has coincided with reports that Gemini 3.5 Pro has been delayed because it failed to meet Google's internal coding performance targets.

The model had originally been scheduled for release in June. According to Bloomberg and other media outlets, Gemini 3.5 Pro fell short of internal benchmarks in code generation and debugging, potentially delaying its launch by several months. Google reportedly attempted to improve performance by fine-tuning the model with additional coding datasets in June 2026, but the results were insufficient, forcing engineers to restart training.

Pressure on Google has intensified following OpenAI's release of GPT-5.6, which significantly improved coding capabilities and cost efficiency, as well as new flagship models from Meta and SpaceXAI — Muse Spark 1.1 and Grok 4.5, respectively — emphasizing coding and agentic AI.

OpenAI, Anthropic and Google have generally been regarded as the three leading frontier AI developers. Analysts say that if Google's next-generation model fails to outperform competing systems from Meta and SpaceXAI, the company could lose its position among the industry's leading frontier model providers.

With Chinese competitors rapidly improving, Google appears to have opted to prioritize model quality over speed of release. Following Bloomberg's report, Google's share price fell nearly 4%. A Google spokesperson said the company continues to release models rapidly while balancing performance and cost efficiency and is testing Gemini 3.5 Pro and an upgraded Flash model with partners.

Chinese AI models gain ground in the US

Meanwhile, data from AI model aggregation platform OpenRouter indicate that Chinese AI models accounted for a record 58% of AI tokens used by US companies during the first week of July, surpassing US-developed models for the first time.

Bloomberg's analysis of OpenRouter's public traffic data found that Chinese models accounted for 58% of total usage, reaching as high as 63% during some weeks. Chinese models processed roughly 18 trillion tokens during the week, compared with approximately 5.5 trillion tokens for US models.

DeepSeek remained the most widely used individual model. Bloomberg and other international media noted that Chinese models accounted for less than 10% of usage at the beginning of 2025, highlighting the rapid shift over roughly a year.

OpenRouter provides unified access to more than 400 AI models from companies including OpenAI, Anthropic, Google, Meta, DeepSeek and Alibaba through a single API. Because its publicly available traffic data reflect actual model usage by AI startups and developers, OpenRouter has become a closely watched indicator of real-world AI market share.

International media attributed the shift largely to aggressive pricing. Developers building coding automation and agentic workflows are increasingly adopting Chinese open-weight models, which can cost 10 to 50 times less than leading US frontier models.