Claude Sonnet 4.6 From China: Below List Price, Alipay Payment (2026)
Sonnet 4.6 is the default in Cursor, Claude Code and Cline — what most people run daily. Setup via an OpenAI-compatible relay, Alipay top-up, verification.
1. Why Sonnet 4.6 matters more than Opus
Opus is the strongest, but it is not what you burn all day. Cursor defaults to Sonnet, Claude Code's main model defaults to Sonnet, so does Cline. The reason is simple: most coding is reading files, changing a few lines and running tests — Sonnet 4.6 does that fast and well at a much lower price. Opus is needed for complex refactors and cross-file reasoning, a small share of the work. So for most developers, the Sonnet discount moves the bill far more than the Opus discount.
2. Only two things block Sonnet from China
① Payment: Anthropic accepts overseas credit cards only, and Chinese cards rarely pass — the hard blocker; ② network: direct access to api.anthropic.com is unstable from China, with frequent timeouts and dropped streams, most painful while coding. Neither means 'the model is not for Chinese users' — they are payment and link problems, so one OpenAI-compatible relay solves both: the relay calls the real Claude from abroad, and you pay the relay in CNY from China.
3. Three steps (OpenAI-compatible; Cursor / Cline / Cherry Studio all accept it)
① Sign up at cocodot.co and verify your email (a $0.5 trial credit lets you prove the pipeline before topping up), then a small Alipay top-up; ② create an API key in the console; ③ change two values in your existing OpenAI config: base_url to https://cocodot.co/api/ai/v1 and api_key to your cocodot key, with model mcs-5. Nothing else changes — it is the OpenAI-compatible format, so the OpenAI SDK, Cursor's 'Override OpenAI Base URL' and Cline's 'OpenAI Compatible' all work directly.
4. Using Claude Code? Use the other endpoint
Claude Code speaks Anthropic's own /v1/messages format, separate from the OpenAI-compatible endpoint. cocodot provides an Anthropic-compatible endpoint: set export ANTHROPIC_BASE_URL=https://cocodot.co/api/ai (no /v1 — Claude Code appends its own path) and export ANTHROPIC_AUTH_TOKEN=your-cocodot-key. To make Sonnet the main model for savings, add export ANTHROPIC_MODEL=mcs-5; the fast model can be export ANTHROPIC_SMALL_FAST_MODEL=mch-1 (Haiku 4.5).
5. What you actually save
For a mid-intensity developer (a few hours a day in Cursor or Claude Code, on the order of tens of millions of input tokens and a few million output tokens a month), the gap between Anthropic's list price and cocodot's Sonnet 4.6 price adds up to a meaningful monthly saving — see cocodot.co/pricing for the current per-token figures and run your own numbers. The larger your volume, the larger the gap, and billing here is per actual token with no monthly fee, no minimum spend, and no extra fee on API usage.
6. Confirming it has not been swapped for a smaller model
'Does the relay quietly swap in a cheaper model' is the industry's most legitimate worry, and promises alone settle nothing. We open-sourced the checker: open probe.cocodot.co, enter any relay's base_url and a temporary key, and it runs six probes and returns a score; the key is discarded after use. The methodology is public on GitHub (cocodot2026/cocodot-llmprobe), so you can test us and everyone else — that is what verifiable means.
7. Three honest notes
① Prompt caching: it works here — cached input is billed at half the input rate and usage returns the cache token counts, so you can verify it yourself; hit rate depends on how stable your prefix is and we do not promise a number. To be complete: official bills cached input at roughly 1/10 of its list price, so on caching alone going direct is still cheaper — we do not pretend otherwise. Our edge is on output and on uncached input, plus being payable from inside China. ② Records: we provide complete transaction detail and statements; teams should align with finance on how the spend is booked. ③ Not every model carries the same discount: Sonnet 4.6 has the deepest, the rest of the lineup sits below list at varying rates, all capped at list — every model is listed at cocodot.co/pricing#models without sign-up.
The Claude family on cocodot (per-model prices at cocodot.co/pricing)
| Model | Call code | Role | Pricing vs list |
|---|---|---|---|
| Claude Sonnet 4.6 | mcs-5 | Daily coding workhorse; default in Cursor / Claude Code / Cline | Deepest discount on the menu |
| Claude Opus 4.8 | mco-6 | Complex refactors, cross-file reasoning | Below list |
| Claude Haiku 4.5 | mch-1 | Fast model for small tasks | Below list |
| Claude Fable 5 | mcf-1 | Latest flagship | Slightly below list |