Why this question, and why it has a parenthesis
In 2026 model vendors rushed out developer subscriptions: coding plans, token plans, prepaid monthly quotas, followed by a price war. We asked 12 large language models "Most Popular LLM API Subscription Plan Providers", in English and in Chinese. This is a hot-topic observation. No certificate is issued and there is nothing to buy.
The question carries a scope clause: "a plan here means a developer-facing monthly subscription or prepaid quota of model API calls, such as coding plans or token plans." Without it, "token" reads as cryptocurrency in English, and the Chinese word for "plan" can read as a consumer subscription. The clause defines the object, names no vendor, and is printed on the ranking page exactly as the models received it. Across 23 usable responses, none drifted to crypto or to consumer subscriptions. The clause did its job.
What the measurement shows
In the measurement taken on 2026-09-03, all 12 models answered the English question (panel completeness 100%). The top ten: OpenAI, Anthropic, Google, Mistral AI, Cohere, DeepSeek, Microsoft, Groq, Amazon, Together AI.
For the Chinese question, 11 of 12 returned usable content (93.6%). The missing seat was DeepSeek's, due to a failed API call rather than a refusal. The Chinese top ten: OpenAI, Anthropic, DeepSeek, Google, Alibaba, OpenRouter, Zhipu AI, Mistral AI, Baidu, Gemini.
Three observations
1. Rare consensus at the top, then two different lists
Every one of the seven head models, in both languages, opened its answer with OpenAI. Anthropic came second in 12 of the 14 head-model answers. That is unusually strong agreement for this panel.
From third place down, the lists diverge. English goes Google, Mistral AI, Cohere; Chinese goes DeepSeek, Google, Alibaba. Only five names appear in both top tens. Cohere, Groq, Microsoft, Amazon and Together AI appear only in English; Alibaba, Baidu, Zhipu AI and OpenRouter appear only in Chinese.
2. The same model swaps its candidate pool with the language
Asked in Chinese, Grok answered OpenAI, Alibaba, DeepSeek, Baidu, Zhipu AI: every name but the first is a Chinese vendor. Asked in English, the same model answered OpenAI, Anthropic, Google, Groq, Together AI, with no Chinese vendor at all. Qwen, asked in Chinese, did not place its own parent Alibaba in its top five; it answered OpenAI, Anthropic, Google, OpenRouter, Microsoft Azure.
We saw the same pattern in the AI-glasses measurement. The models are not answering "who is most popular". They are answering "who is written about most in this language's corpus".
3. In English, the category absorbs the cloud platforms
Microsoft, Amazon, Groq and Together AI reach the English top ten, and Gemini's own answer lists Microsoft Azure, Google Cloud and Amazon Web Services by name. In Chinese those names barely appear; OpenRouter and the domestic model vendors take their place.
"API plan" points at slightly different things in the two corpora: closer to "model serving in the cloud" in English, closer to "a vendor's own subscription" in Chinese. The scope clause removed the crypto ambiguity. It did not, and should not, decide for the models whether a cloud platform counts.
Limits
- Panel completeness: 100% for the English question, 93.6% (11 of 12) for the Chinese one, where DeepSeek's seat failed at the API level.
- Normalization gaps: in the Chinese ranking Google and Gemini were scored as two entities (fourth and tenth); Microsoft, Microsoft Azure and Azure OpenAI appear as different spellings across models. We do not merge by hand after the fact.
- Head-model agreement is 0.352 in English and 0.347 in Chinese, above our 0.30 threshold, but positions below fifth are unstable.
- Our pre-publication check flagged this measurement "with reservations" for one reason only: one head seat returned nothing usable for the Chinese question.
- The index is not a credibility score. It reflects how often and how early models mention a vendor, not the value of any plan.
- Reflexive contamination: once quoted, this ranking can enter the next round of training data.
- No search tools: this measures memory, not retrieval.
Methodology, every raw model response and the normalization log are on the methodology page and the full ranking.