A Chinese open model at 1/40 of Opus price is cracking US lab pricing power —

On August 3rd Alibaba did something unusual: it released a near-frontier model with the weights attached. Qwen3.8-Max, a 2.4-trillion-parameter system that Alibaba says trails only Anthropic's Claude Fable 5, costs $2 per million input tokens and $6 per million output tokens on Alibaba Cloud's API. Anthropic prices its Opus 5 at $5 and $25, and its Fable 5 at $10 and $50. Hong Kong-listed Alibaba rose 7 per cent on the day, to HK$125.20. Investors read a national champion reaching the frontier. The arithmetic reads a price cut wearing a flag.

The headline gap, moreover, understates the scale of what is happening. At the cheap, self-servable end of the Chinese ecosystem the discount reaches roughly one-fortieth of Opus's price. DeepSeek's V4 Flash, the open-weight workhorse, charges $0.14 per million input tokens and $0.28 per million output tokens — inputs around 36 times cheaper than Opus's $5, outputs around ninety times cheaper than its $25. OpenRouter puts Chinese open-weight models 60–90 per cent below leading American APIs; Vercel's production index shows open-weight tokens at about a tenth of the average price on its platform. "One-fortieth of Opus" is not a point but a curve: a quarter at the flagship API, an order of magnitude further where most tokens are actually burned.

The mechanism matters because it decides everything downstream. When a frontier-class model is downloadable, the price of a token stops being a rent that a closed lab extracts and becomes a competitive floor set by whoever can serve the weights most cheaply. American labs have been pricing as if that floor exists. Token prices for leading American lab models fell by almost a quarter between mid-July and mid-August, according to Silicon Data's index. OpenAI cut its lightweight Luna model by 80 per cent, from $1 to $0.20 input and from $6 to $1.20 output, and trimmed its mid-tier Terra by a fifth; its boss, Sam Altman, says he would be "happy to deliver at one-quarter of the price" if competition demands. Anthropic launched Opus 5 at half the price of its own Fable 5, insisting the gap is an internal product decision, "no connection to competitors".

The buyers moved before the sellers did. The share of US-originating tokens flowing to Chinese models on OpenRouter has exceeded 30 per cent every week since early February, peaking at 46 per cent, up from 4.5 per cent in the first half of 2025. DoorDash and Airbnb have adopted cut-price Chinese models to rein in bills; Lindy, an agentic startup, moved all of its traffic from Claude to DeepSeek and says the switch saved millions within months. This is cross-border, dollar-denominated, production traffic. It is not a skirmish inside China's firewall.

Which is not to say American pricing power has vanished. It has moved, not disappeared. Open-weight models processed 29 per cent of the tokens in Vercel's production gateway in June but under 4 per cent of the spending; Anthropic collected 61 per cent of spending while handling 32 per cent of tokens. The jobs people pay for — hard coding, long agentic runs, back-office automation — stay on the closed frontier, where Chinese open weights still trail by an estimated six to nine months, says Kyle Chan of Brookings. Hostinger's AI chief put the strategy crisply: the American labs "have cut the middle and are defending the top." A near-ninefold gap remains between the cost of a given workload on a Chinese open model and on Claude.

Everything then turns on where the middle meets the top, and on what terms. Is the Chinese cost advantage a cost curve or a funded giveaway? Partly engineering: Qwen3.8-Max is a mixture-of-experts model with 95 billion active parameters of 2.4 trillion, and export controls have made Chinese labs ruthless customers of scarce silicon. Partly subsidy. Alibaba's June-quarter net income fell 75 per cent, to 10.4bn yuan, as AI infrastructure spending surged even as cloud revenue grew 45 per cent; its three-year AI-investment plan runs to 380bn yuan. And the tell is that the giveaway stopped being free the moment it reached the frontier. Qwen3.8-Max did not ship under the permissive Apache licence Qwen used before; it carries a bespoke licence requiring any business above a $50m revenue threshold to negotiate separate terms, with model-as-a-service resellers explicitly carved out. Moonshot, maker of the rival Kimi K3, asks large commercial users for up to 30 per cent of revenue. The fully open member of the new family is the smaller, 27-billion-parameter model. "Open weights" increasingly means open at the edge, metered at the frontier.

This is where the two interpretations of the pop in Hong Kong separate. A price war contained in China would show up only in yuan-denominated discounts on domestic clouds, leaving American list prices and Western routing untouched. The opposite has happened. A genuinely global repricing, by contrast, cuts both ways for the celebrant: the giveaway also depresses the price Alibaba Cloud can charge for the inference it sells, which is how Alibaba actually makes money from Qwen. The rally priced the capability and the champion; it did not price the deflation, which is the same event seen from the sellers' side. American labs, moreover, have the keenest of motives to defend what is left of that pricing power. Both Anthropic and OpenAI filed for their initial public offerings in June, with Anthropic's backers reportedly expecting a $2 trillion valuation by October; their S-1 arithmetic depends on revenue per token in a way that no subsidised Chinese balance sheet does.

None of this is settled, which is the point for anyone deciding what it is worth. Four observations would falsify the thesis that cheap Chinese open weights are eroding American pricing power. First, a re-widened capability gap: Qwen's own small flagship still loses on the hardest knowledge-reasoning tests — 30.8 against Anthropic's 40.0 on Humanity's Last Exam — and if the next Opus, Fable or GPT pulls away on those premium tasks, the top tier reclaims its rent. Second, revenue resilience: if American labs' blended revenue per token stops falling while volumes accelerate, they will have monetised the deflation into demand rather than absorbed it in margin. The listed proxies are unbothered so far: AInvest's aggregate signal still labels Alphabet a Buy. Third, a re-floored price: the revenue-sharing and licensing clauses now attached to Chinese open weights are precisely the mechanism that would lift their effective price toward the American mid-tier, shrinking the gap from below. Fourth, compute: export controls that keep Nvidia's best chips out of Chinese labs cap how far engineering frugality can be pushed.

Watch the one number that settles it. The price a leading American lab can charge for a frontier token a year from now, measured against the open-weight floor beneath it, will say whether the crack of 2026 widened into a break or stopped at the commodity layer. The past six months — list prices falling, routing share above 30 per cent, a frontier lead already measured in months — point one way. Pricing power that rests on the right to close a door is a licence, not a physics; August demonstrated that a licence can be copied by better engineers and cheaper electricity. American labs' remaining defence is the lead itself, and the lead is the only thing on their side of the ledger that export controls cannot restore.