The Chinese AI startup Moonshot AI recently unveiled its open-source inference LLM model, Kimi-K2-Thinking, intensifying the AI performance race in a way reminiscent of this year’s earlier DeepSick breakthrough.
Industry sources reported on Thursday that Moonshot AI claimed its Kimi K2 Thinking, released on the 6th of this month, outperformed leading models like OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 across several benchmarks.
At launch, Kimi K2 Thinking scored an impressive 44.9 points on the Human-Level Evaluation (HLE) test, which assesses how closely AI matches human expert performance in specific fields. This score surpassed OpenAI’s GPT-5 (41.7 points), Claude Sonnet 4.5 (32 points), and DeepSick V3.2 (20.3 points).
In the BrowseComp web search capability test, Kimi K2 Thinking achieved a score of 60.2, outperforming GPT-5 (54.9), Claude Sonnet 4.5 (24.1), and V3.2 (40.1). However, it fell short of GPT-5 and Sonnet 4.5 in the SWE benchmark for coding abilities.
CNBC highlighted the model’s cost efficiency, estimating its training cost at just 4.6 million USD, though Moonshot AI denied that this figure was official.
Industry experts see this as another confirmation of the rapid advancement of Chinese AI models, following DeepSick and Alibaba Group’s Qwen Series.
OpenAI responded swiftly, unveiling GPT-5.1, an upgraded version of GPT-5, just a week after Kimi K2 Thinking’s release. The company reported significant performance improvements in math and coding benchmarks.
Additionally, OpenAI introduced a group chat feature for collaborative ChatGPT interactions, piloting it in countries including South Korea, Japan, Taiwan, and New Zealand.
Elon Musk’s xAI entered the fray by releasing Grok 4.1. on Monday.
Grok 4.1 comprises two models: Grok 4.1 (codename: tensor) for immediate responses, and Grok 4.1 Thinking (quasarflux) for deep thinking. Both models briefly topped several benchmarks, outperforming competitors like OpenAI, Anthropic, and Google.
Google quickly countered by launching its next-generation AI model, Gemini3, touting world-class performance. This latest version comes about eight months after Gemini 2.5’s debut.
Gemini3 scored 37.4 points on the Humanity’s Last Exam (HLE) benchmark, surpassing GPT-5.1 and Claude Sonnet 4.5.
Google plans to release Gemini 3 DeepSync, a research-focused extended version, soon.
U.S. analysts suggest that Gemini 3’s launch quickly overshadowed Grok 4.1’s spotlight.
Meanwhile, the performance gap between U.S. and Chinese AI models is rapidly narrowing. Stanford University’s Human-Centered AI Institute (HAI) reports that the gap between top models from China and the U.S. shrank from 103 points in January 2024 to just 23 points in February. The MMLU benchmark gap is expected to plummet from 20 percentage points in 2023 to just 0.3 percentage points by the end of 2024.
In a groundbreaking revelation, DeepSick disclosed on Tuesday in a peer-reviewed Nature paper that the training cost for R1 was only 294,000 USD. This represents only 0.3% of OpenAI’s reported 100 million+ USD in training costs for its foundational model in 2023.