Bull Bear Market
Bloomberg Charted the US–China AI Gap: What the Viral Chart Shows, and What the Numbers Behind It Say
A viral post summarised Bloomberg's chart of the US–China AI race: Kimi K3 near Claude Fable at 70% less, an Anthropic estimate of six to 12 months, GLM-5.3 missing. We check each claim against the sources and other trackers, and ask what cheaper near-frontier models mean for the AI trade.
Akmal Alif · 10 October 2026 MYT
On 20 August 2026 the AI engineer Hesamation shared Bloomberg's new chart of the US–China AI race with a short summary that spread quickly (Hesamation, 2026). It made six points:
China is closing the gap fast.
Moonshot's Kimi K3 is close to Anthropic's Claude Fable and costs about 70% less.
Anthropic had estimated China was six to 12 months behind.
China now has several frontier-level labs.
Zhipu's GLM-5.3 is not yet in the chart, which would favour China further.
US labs still hold the frontier but struggle to charge frontier prices against Chinese competition.
We checked each claim. Bloomberg's chart is copyrighted and is not reproduced here; we describe what it reports.
What Bloomberg measured
Bloomberg's analysis, by Luz Ding, Spe Chen and Hayley Warren, ranks frontier models from both countries on benchmark performance and on cost per task (Ding et al., 2026). Its headline finding is that the US lead is narrowing rapidly. It names Kimi K3 as the standout: performance near Anthropic's most advanced model at a fraction of the cost. Semafor's summary is more balanced. The top American systems still lead on capability, but Chinese labs are winning customers worldwide on price (Capelouto, 2026).
Checking the claims
Claim | What the sources show |
|---|---|
Kimi K3 is close to Fable | Close, not equal. At K3's launch in July, Artificial Analysis's Intelligence Index scored it 57, against 60 for Claude Fable 5 and 59 for GPT-5.6 Sol (Bastian, 2026). |
About 70% cheaper | True of list prices. K3 charged US$3 per million input tokens and US$15 per million output tokens, against Fable 5's US$10 and US$50: exactly 70% lower (Bastian, 2026). Cost per task depends on how many tokens each model uses, so real savings vary. |
Anthropic estimated six to 12 months | Partly. An Anthropic executive, speaking anonymously to Bloomberg days before K3's release, put Claude about six to 12 months ahead (Ghaffary & Metz, 2026). Anthropic's own May 2026 paper says only that US systems are at least several months ahead (Anthropic, 2026). |
GLM-5.3 is not in the chart | Plausible but unconfirmed. GLM-5.3 launched through Zhipu's API on 14 August 2026, and its open weights followed on 29 August (Gigazine, 2026). We could not check which models the chart includes. |
US labs can't charge frontier prices | Supported, with a twist. Semafor reports Chinese labs winning on price. But The Decoder notes that K3 is pricier than any earlier Chinese frontier model, signalling the end of "super cheap" Chinese AI (Bastian, 2026). |
The "several frontier labs" claim also holds up in the sources we read. Moonshot (Kimi), Zhipu (GLM) and DeepSeek all appear in 2026 coverage of models near the US frontier.
Other ways to measure the gap
Bloomberg's is one ruler among several:
Epoch AI. Using its Epoch Capabilities Index, Epoch found that Chinese models lagged the US frontier by an average of seven months from 2023, with a range of four to 14 months (Emberson, 2026).
Bloomberg Intelligence. In October 2026, Bloomberg Intelligence analyst Robert Lea put the benchmark gap at 3%, after DeepSeek's V4.1 Flash release in September. That compared with about 9% in May and about 15% earlier in the year (Rai, 2026).
The trackers measure different things, so their numbers don't add up neatly. They all point the same way, though: the gap is measured in months and single-digit percentages, and it is shrinking.
What it means for the Tech Bull Market
Our reading, for readers following the AI trade:
Frontier pricing is under pressure. When a model within a few index points of the best sells for about a third of the price per token, it is hard to keep charging premium prices for the last few points. That matters most for labs whose valuations assume premium pricing.
The value shifts to what surrounds the model. That means distribution, agents and tools, enterprise trust, and compute. It also means hardware, much of which is made in Southeast Asia, as our post on the World Bank's April 2026 update shows.
Trust is part of the price. Cheap capability is only half the decision for buyers. Zhipu, the maker of GLM, spent September repairing trust after its ZCode app was found packing up users' workspaces, Git history included, as our ZCode post explains. A US tool, SpaceXAI's Grok Build, had a similar episode in July. Buyers now weigh behaviour, not just benchmarks.
References
Anthropic. (2026, May 14). 2028: Two scenarios for global AI leadership. https://www.anthropic.com/research/2028-ai-leadership
Bastian, M. (2026, July 16). Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI. The Decoder. https://the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5-while-signaling-the-end-of-super-cheap-chinese-ai/
Capelouto, J. D. (2026, August 20). China closes the gap in AI race with US. Semafor. https://www.semafor.com/article/08/20/2026/china-closes-the-gap-in-ai-race-with-us
Ding, L., Chen, S., & Warren, H. (2026, August). US lead in the AI race with China is rapidly narrowing. Bloomberg. https://www.bloomberg.com/graphics/2026-us-china-ai-race/
Emberson, L. (2026, January 2). Chinese AI models have lagged the US frontier by 7 months on average since 2023. Epoch AI. https://epoch.ai/data-insights/us-vs-china-eci
Ghaffary, S., & Metz, R. (2026, July 18). Moonshot's Kimi upends conventional wisdom on US lead over China. Bloomberg News (via Bloomberg Law). https://news.bloomberglaw.com/artificial-intelligence/moonshots-kimi-upends-conventional-wisdom-on-us-lead-over-china
Gigazine. (2026, August 29). The Chinese-made AI model 'GLM-5.3' has been released as an open model free of charge as promised, a high-performance model comparable to Claude Fable 5 and GPT-5.6 Sol. https://gigazine.net/gsc_news/en/20260829-glm-5-3-open/
Hesamation [@Hesamation]. (2026, August 20). [Post summarising Bloomberg's chart of the US–China AI gap] [Image attached]. X. https://x.com/Hesamation/status/2090356790709887061
Rai, S. (2026, October 4). US lead in AI over China narrows after DeepSeek gains, BI says. Bloomberg News (via Bloomberg Law). https://news.bloomberglaw.com/artificial-intelligence/us-lead-in-ai-over-china-narrows-after-deepseek-gains-bi-says