Checked Nov 7, 2026Start a new check

What was checked

New open-source AI model from China outperforms GPT-5 on all reasoning benchmarks and is completely uncensored

Assessment summary

Needs caution

Kimi K2 Thinking is described as open source and reported to outperform GPT-5 on selected benchmarks, including Humanity’s Last Exam, BrowseComp, and Seal-0. The evidence does not show that it beats GPT-5 on all reasoning benchmarks, and coding performance was only comparable. The available source does not establish complete absence of censorship for Kimi K2 Thinking. A NIST evaluation found censorship in other Chinese models, including downloaded open-weight models, so the blanket claim requires model-specific evidence.

Key findings

Model assessments

Assessment summary

  • The claim appears to refer to Kimi K2 Thinking, which its developer says outperforms GPT-5 on several key benchmarks, not all reasoning benchmarks zdnet.com.
  • Its open-source status is supported, but “completely uncensored” is not established; authoritative evidence shows censorship in at least some Chinese models nist.gov.

Claims reviewed

A new open-source AI model from China outperforms GPT-5 on all reasoning benchmarks

Needs context
Kimi K2 Thinking is described as open source and reported to outperform GPT-5 on selected benchmarks, including Humanity’s Last Exam, BrowseComp, and Seal-0 zdnet.com. The evidence does not show that it beats GPT-5 on all reasoning benchmarks, and coding performance was only comparable zdnet.com.

This claim may mislead readers without added context.

The model is completely uncensored

Can’t verify
The available source does not establish complete absence of censorship for Kimi K2 Thinking. A NIST evaluation found censorship in other Chinese models, including downloaded open-weight models, so the blanket claim requires model-specific evidence nist.gov.

Assessment summary

  • The claim contains partially accurate elements but makes absolute statements that are misleading or unsupported.
  • While Chinese open-source models have indeed outperformed GPT-5 on some specific benchmarks, the claim that one model outperforms GPT-5 "on all reasoning benchmarks" is not substantiated by available evidence.
  • Additionally, the "completely uncensored" claim is problematic—Chinese models typically include censorship by default, though some can be modified.

Claims reviewed

New open-source AI model from China outperforms GPT-5 on all reasoning benchmarks

Needs context
While Chinese models like DeepSeek and Kimi K2 have outperformed GPT-5 on some specific reasoning benchmarks, evidence does not support outperformance on ALL reasoning benchmarks. A 2026 USCC report noted that no single Chinese model is better than GPT-5.4 overall as of April 2026. Chinese models excel in specific areas (math, coding) but not universally across all reasoning tasks.

This claim may mislead readers without added context.

The model is completely uncensored

Needs context
Chinese AI models like DeepSeek, Qwen, and Kimi K2 include built-in censorship by default. While some third-party companies have created 'uncensored' versions by removing censorship filters, the original Chinese models are not 'completely uncensored.' The claim conflates open-source availability with lack of content filtering.

This claim may mislead readers without added context.

Assessment summary

Claims reviewed

A new open-source AI model from China outperforms GPT-5 on all reasoning benchmarks

Needs context
Press coverage documents Chinese open-source models (e.g., Moonshot's Kimi K2, GLM variants) outperforming GPT-5 on several reasoning and coding benchmarks content.techgig.com aidailypost.com. However, sources report wins on specific benchmark suites or metrics, not a definitive sweep of every reasoning benchmark, so the absolute wording "all" is unsupported content.techgig.com atomic.chat.

This claim may mislead readers without added context.

The model is completely uncensored

Can’t verify
Public reports describe the models as open-weight or less restricted, but they do not provide independent verification that there are literally no content or safety filters in production deployments; companies often apply moderation layers outside model weights content.techgig.com aidailypost.com.

Sources cited

17 sources

Share this result