Despite the Great Firewall and extra costs, Chinese developers are going to great lengths to use OpenAI’s GPT-5.6. As reported by the South China Morning Post, users in mainland China – where the service is officially blocked and reachable only via VPN or third-party proxy – consider the premium well worth paying. That’s no small detail: Chinese frontier models undercut on price, yet the choice falls on an inaccessible and more expensive product. The reason is a cold math where cost per token stops being the only variable.

The calculation that leads to climbing the firewall is not purely technical: it reflects a performance gap that savings alone cannot close. Those building complex applications – code generation, document analysis, multi-step reasoning – know that a less capable model produces errors that cost time, revisions, and lost trust. When output quality affects productivity, every cent saved on a single token translates into a far larger hidden cost. Thus VPNs, proxies, and tariff markups become acceptable items in a TCO where the dominant factor is the model’s ability to get the job done.

This behavior signals something structural for the AI ecosystem. China has world-class labs and competitive models, but the fact that a niche of developers is willing to pay more for a blocked service suggests the capability gap has not yet closed. It is not a question of lacking alternatives, but of asymmetry in the most demanding use cases. For Chinese teams, this could further accelerate R&D investments to bridge the gap, or push toward distillation techniques and training on frontier model outputs, lengthening the chain of technological dependence.

From an on-premise deployment perspective, the episode has an immediate echo. If access to the most powerful models depends on blocked servers or intermediaries, data sovereignty and operational continuity become fragile. A self-hosted infrastructure based on competitive open models – even if it currently requires significant investment in hardware and tuning – offers a path to break free from an ecosystem where quality is hostage to geographic availability. It is no coincidence that the open-weight LLM market is attracting growing attention: the ability to run inference locally, without negotiating with firewalls and proxies, restores control and cost predictability.

The paradox of GPT-5.6 in China ultimately shows that the competitive barrier is not just token price, but the full range of capabilities a model provides. And as long as that gap remains, there will be those willing to surmount it by any means.