Bloomberg: US-China Top AI Model Benchmark Gap Narrows to 3%, a Record Low

How the 3% Was Calculated

Bloomberg Intelligence senior analyst Robert Lea wrote in an October 5 report: after DeepSeek’s September release of V4.1 Flash, the benchmark gap between China’s top models and their US rivals narrowed to about 3%. The figure was 9% in May, 15% at the start of the year.

The gap comes from LiveBench, a public leaderboard scoring models on Q&A, puzzles, and hands-on tasks — IQ-test style. On September’s board, V4.1 Flash ranked 6th globally. It’s the highest rank a Chinese model has taken since DeepSeek broke out with the R1 reasoning model in 2025. DeepSeek scored 81.1; Anthropic’s best hit 83.4. Lea’s verdict: now comparable to top systems from Anthropic and OpenAI.

The Gap Isn’t in Averages — It’s in Lineups

Close scores don’t mean caught up overall. Only 3 of LiveBench’s top 15 are Chinese models; US vendors still dominate in sheer count of top-tier models. The gap is also uneven by category: per 动区’s check of the raw leaderboard data, V4.1 Flash topped Agentic Coding (a test where models edit code, run tests, and finish tasks themselves) with 77.3 — the best on the board, versus 66.1 for Claude Fable 5.1 — but trailed in reasoning (86.7 vs 91.7) and language (81.2 vs 89.5).

A Rival Tens of Times Cheaper

Bloomberg attributes it to two things: deepening Chinese AI research capability, and skill at optimizing models for domestic hardware. Lea bluntly questioned “how much longer US technological dominance in AI can last,” and the effectiveness of export controls.

Then there’s price. DeepSeek’s list pricing: V4.1 Flash output at $0.60 per million tokens off-peak, $1.20 peak. Anthropic’s: Claude Fable 5.1 output at $50 per million tokens. A 42–83x unit-price gap. But unit price isn’t total task cost — different models burn different token counts for the same job, so discount that multiple.

Rankings Shift; Making Money Stays Hard

Lea flagged two things: rankings will keep moving; and however high the rank, commercialization can still be brutal. China now has 1,100+ LLMs in a fierce price war — the industry may not turn profitable until 2030 at low margins. On consumer apps, ByteDance’s Doubao leads; DeepSeek’s and Tencent’s chatbots are still free.

The timing is sensitive. Anthropic and OpenAI are preparing IPOs at trillion-dollar-scale valuations, with capability leads as the core valuation argument. Chinese models are closing the user-count gap on price; mentions of open models on US earnings calls are already 6x last year’s. More companies are mixing and matching: pricey top models for critical tasks, cheap open models for daily work.