Invest1 publisher2 min readPublished
US AI labs keep a six-to-eight-month lead in reasoning and cyber tasks, but cheaper Chinese models gain market share
Chinese models handled 50% to 67% of OpenRouter's token traffic by mid-2026, with DeepSeek's V4-Pro priced near $3.96 per million output tokens. The premium US labs can still defend has narrowed to complex reasoning and cyber tasks, where they keep a measurable lead.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- DeepSeek followed its January 2025 R1 reasoning model with the open-weight V4 series, which began rolling out in April 2026.
- A May 2026 CAISI evaluation found the best Chinese models trail US frontier models by six to eight months in complex reasoning and cyber tasks, with V4 Pro about eight behind.
- The US holds about 74% of compute power and leads in high-impact patents, while China publishes more papers and leads on some citation metrics.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost A developer pushing a billion output tokens through V4-Pro pays about $3,960. A US lab charging more has to justify the difference on work where its lead can be measured.
- contradiction Stanford's 2.7% gap and CAISI's months-long lag measure different work, so a buyer of general tasks sees near-parity while a buyer of reasoning or cyber work still sees a US lead.
- decision OpenRouter lets developers pick a model for each request, so US labs have to win general-purpose requests on price, the ground Chinese models have already taken.
The 2.7% level is easier to rely on than the size of the fall. Taken together, Stanford's two figures put the March 2026 gap at no more than 3% of the May 2023 one, closed over 34 months [4] [5]. The two numbers do not fit well on one scale, though. If 39 points equals 2.7%, the leading score is near 1,440 points [1], and a gap of more than 1,300 points would be about nine-tenths of that entire score [3].
Bloomberg, as cited by Crypto Briefing, credits DeepSeek's progress with much of the narrowing [11]. Price is the other half of the case. Crypto Briefing's report says V4-Pro's price significantly undercuts US competitors [4]. It does not include a US price, a split of spending by task, or any revenue, margin or valuation figure for US labs or the cloud companies behind them.
The OpenRouter share needs one adjustment before it says anything about money. The report takes it to mean many developers have already chosen cheaper Chinese alternatives [9]. If those models cost less per token, their share of developer spending on the platform is below their share of tokens. The distance between those two percentages measures the premium developers still pay US labs. OpenRouter routes requests across many models [5], so its users are the developers for whom switching costs least.
From here, US labs can hold a premium on complex reasoning and cyber work, where OpenAI and Anthropic still lead [6], and give up general traffic. The specialized lag could instead close the way the general one did; DeepSeek went from its R1 model in January 2025 to an open-weight V4 series in April 2026 [3]. Or the US compute lead [8], built on the hardware that export controls were designed to keep from Chinese labs [10], lets US labs cut their own prices and keep volume at a lower margin.
I think the first path is the one this record supports. Applied to the valuations of US labs and their cloud backers, it supports one claim: price per token is under pressure in general-purpose work, and the premium that remains depends on a lead of months on a subset of tasks [7]. Against that, eight months is a long time if most spending sits in hard tasks. The view is wrong if US labs keep charging more for general work, where the measured gap is 2.7% [1], and still hold their volume. Buyers would then be paying for something the benchmarks do not measure.
What to watch
- CAISI's next evaluation: a reasoning and cyber lag below six months would remove the main measured basis for a US price premium.
- OpenRouter share data: Chinese models moving above the 67% top of the mid-2026 range would show volume still leaving US labs.
- Any US lab price cut on general-purpose models toward V4-Pro's $3.96 per million output tokens.