Chinese developers are climbing the firewall to pay more for GPT-5.6, and doing the maths on why

TL;DR

OpenAI’s GPT-5.6 release (Sol, Terra, Luna) is winning praise from users in mainland China who reach the blocked service via VPNs, despite costing several times more per token than Zhipu’s GLM-5.2 or DeepSeek V4. Their argument is that per-token pricing is the wrong metric: Artificial Analysis found GPT-5.6 finished coding-agent tasks using roughly a ninth of the output tokens DeepSeek V4 needed, while scoring higher. Efficiency, not the rate card, is what lands on the invoice.

OpenAI’s new GPT-5.6 models are winning praise from an unlikely group. Users in mainland China, where the service is blocked and reachable only by VPN or third-party proxy, told the South China Morning Post the models are worth the extra cost.

That is a striking verdict, because the extra cost is not small. Chinese frontier models undercut OpenAI by a wide margin.

The release, which landed after US government approval, brought three tiers. There is the flagship Sol, the balanced Terra, and the lightweight Luna.

The price gap is real

Sol costs $5 per million input tokens and $30 per million output tokens. Terra runs at $2.50 and $15, while Luna sits at $1 and $6.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

OpenAI says Terra roughly matches GPT-5.5 at about half the price. Even so, the flagship and mid-tier models remain well above what Chinese labs charge.

Zhipu’s GLM-5.2 is priced around $1.40 per million input tokens and $4.40 for output. DeepSeek V4 goes lower still, at up to $0.44 and $0.87.

On paper, that makes DeepSeek roughly a tenth the price of Sol on output. The price war DeepSeek touched off has left OpenAI looking expensive in Chinese terms.

Why buyers pay anyway

The counter-argument is that per-token pricing measures the wrong thing. What matters is how many tokens a model burns to finish the job.

By that measure the gap flips. According to Artificial Analysis, GPT-5.6 completed coding-agent tasks using roughly one ninth of the output tokens DeepSeek V4 consumed, while scoring higher overall.

Fewer tokens means lower inference cost and faster execution. A model that is ten times the price per token but uses a ninth as many is not obviously the expensive option.

That logic is now central to enterprise buying, where falling token prices have coincided with rising bills. Efficiency, not the rate card, is what shows up on the invoice.

What users say they notice

The reported gains are about staying on task. Li Yitao, co-founder of Canada-based AI start-up Quotaflow, said ChatGPT stands out for solving problems methodically, making it suited to large projects.

Li argued enterprise customers now want systems that can make professional judgments and sustain long, multi-step reasoning. General-purpose answering is no longer the point.

Vincent Liu, a user in Hubei province, said the model drifts less than GPT-5.5 during multi-turn conversations. It can “think through a task and expand based on the original framework” rather than being led off course, he said.

On Xiaohongshu, a user surnamed Yan praised the model’s use of subagents and its “extremely high token efficiency”. A Weibo user reported cutting interface development work from hours to around 30 minutes.

The awkward part

None of this is a market OpenAI can serve. Its products are not officially available on the mainland, so every one of these users is reaching them through a workaround.

OpenAI also launched ChatGPT Work, an agent that gathers information across apps and produces finished documents. It is aimed squarely at the enterprise workflows these users describe.

The praise is real, but it is worth keeping in proportion. These are self-selected enthusiasts willing to jump a firewall, not a representative sample of Chinese developers, most of whom are consolidating onto domestic tools.

Still, the efficiency argument is the one that travels. If a pricier model finishes the work in a fraction of the tokens, the cheapest label on the shelf stops meaning much.