Hangzhou Priced the Legal-AI Discount

A million input tokens on the top-tier Chinese frontier model now runs at roughly twenty-eight cents. That is not a promotional teaser rate; it is a rate a mid-tier law firm can pull down as a monthly subscription, on domestic silicon, at yuan-billed rates. It is also the number that makes contract review, discovery filtering, precedent search and drafting look, on the invoice, like a different business than they did a year ago.

The invoice is where I want to stay for a moment, because on 30 April 2026 a court in Hangzhou put a boundary around what those paper savings can be turned into. The two facts together, the cheap model layer and the ruling on top of it, are the shape of legal-AI economics in China this quarter. Everything else in the trade press is decoration.

The bill you can read on the invoice

Alibaba Cloud's product page for Farui Plus, the Tongyi Farui legal model, is dry reading and useful precisely for that. The page describes what the model does: legal question answering, case analysis, document generation, contract review. The latest update wires the model into Alibaba's agent stack for research retrieval, cross-referencing statutes with precedent and returning briefs with verifiable citations. That is a description of a workflow, not a demo, and it is priced against tokens rather than seat licenses.

The inference layer under it is cheaper than any Western equivalent by a wide margin. DeepSeek's published rate of $0.28 per million input tokens has sat, since spring 2026, roughly 12× below GPT-4o's rate for comparable classes of work. The training-side economics are the other axis. DeepSeek V4 was trained on a 12,000-chip Huawei Ascend 910B cluster with about a 30% lower running cost than an equivalent Nvidia setup, and once the model was trained the marginal cost of serving it dropped again on domestic silicon. Enterprise-wide inference budgets in the region have compressed accordingly, with SCMP tracking a 2026 low across Chinese open-source models. Nothing about that curve is finished. On 4 August 2026 Alibaba shipped Qwen3.8-Max, a 2.4-trillion-parameter model with a 1 million token context window, and put "legal document review" among the workloads it was explicitly positioned for.

That is the compute floor a Chinese law firm now operates against. The industry-level demand signal sits on top of it. The Wolters Kluwer 2026 Future Ready Lawyer Survey, released 10 March 2026, covers 810 lawyers across the US, China and nine European countries. Ninety-two percent report using at least one AI tool in their daily work. Sixty-two percent report saving between six and twenty percent of their weekly working time. The China cohort tracks the broader sample, and Wolters Kluwer's companion note on Chinese firms is explicit that some firms, Hui Ye among them, have built their own AI-based management systems on top of the tooling layer.

Stack the three items and read what the invoice becomes. Cheap tokens. A domain-tuned model. A workforce measurably shifting how much time it spends on the tasks the model does. The unit economics of a document-review batch are being rewritten at the token layer, and the yuan cost sits inside a range a Beijing or Shanghai partner can absorb without a capex conversation.

Shenzhen and the throughput number

Shenzhen Intermediate People's Court has been the productivity data point everyone in Chinese legal-tech circles points to. Each judge handled 744 cases in 2025, up by 249 year over year. Roughly fifty percent more cases per judge. The court attributes the jump to a domain-specific LLM assistant it began building in 2024, the first intelligent judicial assistance system in China running on a custom legal model. The system now covers, on the court's own account, 85 procedures across civil, administrative and criminal work: filing, review, hearings, document preparation.

Read the number the way a working lawyer would. Two hundred and forty-nine more cases per judge is a compressed reading pipeline: retrieval and briefs pre-organised for review, standardised drafting handled at a first pass, extraction of the salient facts from filings a paralegal would previously have marked up by hand. It is not a claim that judges are fifty percent cheaper to employ. The Shenzhen court paid its judges the same wage before and after the tool went in.

That gap is what the Wolters Kluwer numbers already hinted at. Time saved is not the same as headcount removed, and the discipline of holding the two apart is where every honest legal-AI conversation should start. The gap between throughput and payroll is, in a formal sense, an accounting gap. It is also, and this is the piece I care about today, where Chinese labor law has now planted itself.

What Hangzhou fixed

On 30 April 2026 the Hangzhou Intermediate People's Court published a ruling in what the court described, in its own case-selection language, as a "typical example" of protecting AI-enterprise and worker rights (English release, State Council Information Office). The dispute was between a former quality-inspection supervisor, surnamed Zhou, and a Hangzhou fintech firm that had replaced his role with an LLM-based quality review pipeline. The Sixth Tone write-up carried the specific numbers.

Zhou's salary before the dispute was 25,000 yuan per month. The firm proposed reassigning him to a lower-level role at 15,000 yuan. He refused. The firm terminated him, offering roughly 311,695 yuan in compensation and citing organisational restructuring. Zhou sued. The lower court found for Zhou. The intermediate court upheld the finding. Zhou was awarded more than 260,000 yuan.

The reasoning is the piece any legal-AI deployment lead should be reading this month. Judge Shi Guoqiang held that AI-driven cost advantage does not, under Article 40 of China's Labor Contract Law, count as a "major change in objective circumstances." The Article 40 language is the ordinary redundancy provision, the ground on which a firm may lawfully end an employment contract when the world shifts in a way the parties did not foresee: business collapse, unavoidable loss reduction, a major structural change. Cheaper inference from a domain-specific model, the court held, is a business decision the firm made about its own margin. It is not an objective condition to which the firm is responding. The costs of the technological transition, in the court's phrasing, cannot be pushed onto the worker.

The ruling is doing work in the wider labour market four months on. NPR's 10 August 2026 follow-up reports Chinese courts continuing to side with workers on AI-replacement claims while employee anxiety refuses to settle, which is what you would expect: case law is the floor of protection, not the ceiling of felt risk. The Fisher Phillips client alert, a US management-side employment firm advising global employers, reads the ruling as a warning and disagrees with treating it as narrow to China. Their read is worth engaging with rather than dismissing. Fisher Phillips is not wrong that the Hangzhou reasoning has jurisdictional reach: the "objective circumstances" language exists, in some form, in most civil-law employment systems, and a firm running a global AI-cost-reduction playbook would be prudent to assume similar reasoning will surface in Paris, Milan, São Paulo. Where I would push back on their read is the implication that the ruling is a brake on adoption. That reads the wrong noun. What the ruling brakes is how the savings can be booked, not whether the firms buy the model.

The numbers, read together

Here is what a Chinese law firm can now truthfully claim on its operating budget this quarter. Compute cost per contract review pulled down by an order of magnitude against the 2025 line. Turnaround time compressed by the multiples Shenzhen has demonstrated on judicial workflow, call it somewhere between twenty and fifty percent depending on the task. Adoption at, roughly, industry-normal levels for 2026, which per the Wolters Kluwer sample is essentially universal. That is real, and it is bookable.

Here is what the firm cannot now claim, without exposing itself to Zhou-shaped litigation. A headcount reduction traced to AI cost advantage, framed as though the reduction were an objective response to shifted market conditions. A downward salary revision imposed unilaterally on a role the model displaced. A restructuring memo that names AI-driven efficiency as the operative reason without the tenant-level facts of failing revenue or an actual reorganisation of the practice line.

The measure that survives contact with the Hangzhou reasoning is per-contract cost, minus per-lawyer-hour repositioned into higher-margin work the firm can charge the client for. If the firm has not decided what the repositioned hour actually does, whether that is a new client, a new practice line, a deeper matter, or faster turnaround the client is willing to pay a premium for, the discount is a rounding entry on the P&L rather than a business result.

The pilot

In an office tower on Fuchun Road, three blocks from the Hangzhou Intermediate People's Court, a partner opens the Sixth Tone piece on her tablet before Monday's session on the firm's new contract-review pilot. The screen behind her shows the Farui Plus console, logged in and waiting. On the shelf above her desk is a copy of the Labor Contract Law with a yellow tab marking Article 40. She does not close either window. She copies the ruling into the pilot's channel and pins it above the meeting agenda, with three words over the link: read this first. The pilot begins on schedule the following Monday, one seat lighter than it would have been the week before.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.