AI Development

Kimi K3.1 Pro costs half as much as Opus 5.5: where your data goes in Claude Code and how to keep it out of China

Kimi K3.1 Pro costs half as much as Opus 5.5, but Moonshot’s guide sends your code to China. Vercel AI Gateway lets you drop Alibaba and Moonshot and add ZDR.

Luca Di Domenico

Luca Di Domenico

· 6 min read

In short

Moonshot’s Kimi K3.1 Pro costs $2 and $10 per million tokens, half the price of Opus 5.5, and scores 65.7 versus 66.4 on Terminal-Bench 4.0 according to Moonshot. With Moonshot’s official guide, prompts and code go to the Chinese company, and on Alibaba in Frankfurt the Global scope can process data in China too. With Vercel AI Gateway you remove Alibaba and Moonshot from the providers and turn on Zero Data Retention, but for Kimi the only region you can pin is still the United States.

Moonshot’s Kimi K3.1 Pro costs half as much as Claude Opus 5.5 and scores almost the same on Terminal-Bench 4.0: 65.7% versus 66.4%, figures measured by Moonshot. The catch is your data. If you connect it to Claude Code with Moonshot’s official guide, your prompts and the code Claude Code reads go to a Chinese company. And Alibaba’s servers in Frankfurt store your data in Germany, but they can also process it in China. The simplest way around this is Vercel AI Gateway with two settings: remove Alibaba and Moonshot from the allowed providers and turn on Zero Data Retention.

What did Moonshot release on November 4?

On November 4, 2026, Moonshot, the company that builds Kimi, released the Kimi K3.1 Pro weights on Hugging Face. Since then anyone can download the model and install it on their own servers.

Kimi K3.1 Pro is a huge model, though: 2.8 trillion parameters, 104 billion of them active per token, and about 1.56 TB of files for the weights alone. Running it takes very expensive hardware. I explained why local AI for coding is rarely worth it today, and a model this size makes the point even stronger.

So if you want Kimi K3.1 Pro in Claude Code, an API is the practical route. That raises the question few people ask: where does your data end up?

How much does Kimi K3.1 Pro cost compared with Opus 5.5?

Kimi K3.1 Pro costs half as much as Opus 5.5:

  • Kimi K3.1 Pro: $2 per million input tokens and $10 per million output tokens, the providers’ price on Vercel AI Gateway;
  • Claude Opus 5.5: $4 per million input tokens and $20 per million output tokens, from Anthropic’s official price list.

I covered Opus 5.5’s price cut versus Opus 5 in what changes for Claude Code users.

What happens to your data with Moonshot’s official guide?

Moonshot’s official "Use Kimi in Claude Code" guide sets ANTHROPIC_BASE_URL to https://api.moonshot.ai/anthropic. With that setup Claude Code sends everything to Moonshot: what you type in the chat and all the code Claude Code reads from your projects.

Moonshot is the Chinese company that develops Kimi. For a personal project that may be fine. For a client’s or a company’s code, it is a decision to make knowing where the code goes.

Does Alibaba in Frankfurt keep your data in Europe?

No. Alibaba Cloud Model Studio offers Kimi K3.1 Pro from its Frankfurt region too, and it looks like the European option. On Alibaba, though, two separate settings are involved:

  1. The region (Frankfurt, for example) decides where your data is stored.
  2. The deployment scope decides where the model actually processes your requests.

For Kimi K3.1 Pro, Frankfurt only offers the Global scope. Alibaba defines it as "any available node, including Chinese mainland and overseas". So your code is stored in Germany, but it can be processed in China.

How do you use Kimi K3.1 Pro in Claude Code without sending your data to China?

Because the weights are public, non-Chinese companies serve Kimi K3.1 Pro as well, such as Together AI, Fireworks and Baseten. The simplest way to use them is Vercel AI Gateway, which routes requests across providers.

Setup takes three steps:

  1. Run the Claude Code setup command: npx vercel@latest ai-gateway setup.
  2. In Claude Code, pick Kimi K3.1 Pro with the /model command.
  3. In the AI Gateway settings, change two things.

These are the two settings:

  • Remove Alibaba and Moonshot from the allowed providers. It is the "Restrict allowed providers" option in the AI Gateway Settings tab. Only the team owner can change it, it is available on Pro and Enterprise plans, and it costs $0.10 per 1,000 successful requests.
  • Turn on Zero Data Retention for the whole team, in the same Settings tab. It applies to every request from the team, including those sent by an agent like Claude Code.

From then on, requests only go to providers that delete prompts and responses as soon as they finish. If none of them is available, the request fails instead of landing at a provider you have not approved: your data never leaves.

One detail to know before you choose: in Vercel’s provider list, some providers, such as Fireworks and Nebius, serve Kimi K3.1 Pro with weights quantized to FP4. They are lighter to run, and answer quality may differ.

Does Vercel AI Gateway keep your data in Europe?

No, and that is worth saying plainly. Vercel AI Gateway lets you pin the region where the provider processes requests, but for Kimi K3.1 Pro the only region you can pin today is the United States. Without a pinned region, the gateway picks whichever region it prefers.

With this setup your code goes to neither Moonshot nor Alibaba, and providers delete it after each request. It still does not stay in Europe.

Can you trust the Kimi K3.1 Pro benchmarks?

Moonshot measured the benchmark numbers itself, on the model card. Kimi K3.1 Pro was tested with Claude Code as the harness, while the Opus 5.5 scores come from published posts and leaderboards. The two headline results:

  • Terminal-Bench 4.0: Kimi K3.1 Pro 65.7, Opus 5.5 66.4;
  • ProgramBench, another coding test: Kimi K3.1 Pro 31.5, Opus 5.5 65.0.

On Terminal-Bench the two models are nearly tied. On ProgramBench Opus 5.5 scores more than twice as much. Half the price for the same performance holds on one test, not on all of them.

So should you switch to Kimi K3.1 Pro?

Kimi K3.1 Pro is a cheaper alternative to Opus 5.5 for Claude Code users. Before switching, two things matter: choosing who receives your code, and testing the model on your own projects, because the vendor’s benchmarks do not tell you how it handles your code.

If you use Claude Code and want a cheaper model without sending your data to China, the time to set up providers and Zero Data Retention is before your first prompt.

Sources

Want to know what this means for your product?

If you're building software with AI and want a second opinion on how you're working, tell me about your project.

Tell me about your project
Claude CodeOpen-weight modelsData privacyAI Development

Follow the blog

One email for every new article

I only write when an article is out: title, summary and link. You can unsubscribe with one click from any email.

RSS feed

An RSS feed is the up-to-date list of articles: add it to a reader app like Feedly or Inoreader and new articles show up there, no email needed.

lucadidomenico.studio/en/blog/rss.xml