Chinese Students Buy GPT-5 and Claude Access for 97% Off, Experts Warn of a Risky Catch

Chinese students and developers are reportedly accessing frontier AI models including OpenAI's GPT-5 and Anthropic's Claude at discounts of up to 97 per cent through unofficial proxy networks, according to an Oxford China Policy Lab investigation and developer discussions circulating online.

The services are advertised across platforms such as Taobao, Xianyu, Telegram and GitHub, where vendors claim API-level access to frontier models through shared subscriptions routed via intermediary proxy infrastructure.

Researchers describe the phenomenon as a parallel access economy in which official pricing is effectively decoupled from end-user cost through third-party redistribution layers.

Analysts including Zilan Qian describe the emergence of a 'transfer station economy' in which AI API access is redistributed through informal intermediary networks outside official licensing channels.

Verified Listings and Distribution Model

The Oxford China Policy Lab investigation identifies structured grey-market listings offering access to frontier AI models via 'transfer stations' that operate outside authorised provider channels.

These intermediaries process user requests on behalf of multiple subscribers, forwarding them through external servers before reaching official APIs. Listings observed across Chinese second-hand marketplaces and encrypted messaging platforms advertise pooled subscriptions and rerouted endpoints as discounted access alternatives.

This distribution model is based on subscription pooling and request forwarding across rotating international nodes, reducing traceability while inserting third-party control over request handling and data transmission.

Pricing Inversion and Structural Break

Official API access to frontier models such as GPT-5 and Claude is estimated at roughly $50 to $120 per million tokens (£40 to £95), depending on model tier and output complexity.

Proxy-based services circulating in grey markets reportedly offer comparable output for around $1 to $4 per million tokens (£0.80 to £3), enabled by shared subscription usage and distributed account access.

In practical terms, a workload costing around $100 (£80) through official APIs can be processed for approximately $3 (£2.40) via proxy systems, a reduction of roughly 97 per cent. Researchers describe this not as incremental discounting but a structural inversion of frontier AI pricing driven by redistribution rather than efficiency gains.

Developer Usage Patterns

Among computer science students, proxy access is increasingly embedded in everyday development workflows.

A typical pattern described in university environments involves continuous use of GPT-5-class and Claude-class models for debugging, documentation, testing, and code generation. In this setup, AI tools are integrated directly into coding workflows and used at high frequency without cost constraints.

This usage pattern enables sustained high-volume token generation that would be economically prohibitive under official pricing structures.

Evidence Hierarchy: Verified Data and Claims

Confirmed findings show structured grey-market listings across Taobao, Xianyu and Telegram channels advertising shared or rerouted access to frontier AI systems.

User-generated accounts on Reddit and short-form platforms describe extreme usage scenarios involving very high token volumes at minimal cost. These claims remain unverified and are treated as anecdotal rather than representative.

Analysts note that proxy systems introduce an expanded data pathway in which prompts, code, and outputs are processed through third-party infrastructure before reaching official model providers, increasing exposure outside provider-controlled systems.

Social Amplification

Broader online commentary has amplified the discussion, including a YouTube video referencing claims of ultra-low-cost access to GPT- and Claude-class models via unofficial proxy networks and shared subscriptions.

While not evidential, such content reflects growing visibility of grey-market AI access within developer communities and contributes to its perceived normalisation.

Risk: Third-Party Data Exposure

Researchers identify the primary risk not as model performance but data control.

In proxy-mediated workflows, user prompts and code are transmitted through intermediary servers before reaching official APIs. While outputs remain consistent with direct usage, the intermediary layer introduces unknown data handling practices outside provider oversight.

In practical terms, this creates potential exposure of proprietary code, prompts, or sensitive business logic, with no transparency over logging, retention, or reuse by third-party operators.

Market Expansion

The grey-market ecosystem reflects a structural mismatch between frontier AI pricing and sustained high-volume demand.

Subscriptions are frequently pooled or rotated across multiple users, effectively functioning as shared compute pipelines that significantly reduce marginal cost.

For students and early developers, this creates near-unrestricted access to frontier models at minimal incremental expense, enabling workflows that would otherwise be financially unviable.

Regulatory and Platform Response

The practices are widely considered a breach of terms of service by major AI providers, including OpenAI and Anthropic. Analysts expect increased enforcement through anomaly detection, account clustering analysis, and tighter regional controls as proxy usage expands.

Researchers also warn that rerouted traffic may distort usage data used for infrastructure planning, potentially influencing future pricing models and enforcement mechanisms.

Parallel AI Access Economy

The findings point to the emergence of a parallel access economy in which official pricing structures and unofficial distribution systems operate at fundamentally different cost levels.

For users, the benefit is scalability and affordability. For providers, the concern is not only revenue leakage but the fragmentation of data governance as frontier AI access increasingly moves through external intermediaries outside controlled infrastructure.