“Cheap” Claude API via third-party proxies: cheap, but what are you really paying with?

Almost every week, AI groups carry posts selling “cheap Claude keys” or “Claude endpoints 50–80% off” — sometimes 90%. The pitch is tempting for developers: change the base_url in your code and you’re done, no international card, no business verification. This article explains how these services work, why they can be so cheap, and the real price you pay.

How API “transfer stations” work

In Chinese developer communities these services are called transfer stations. The common model:

1. You prepay for a “token balance”.
2. Your application sends requests to the reseller’s endpoint instead of api.anthropic.com.
3. The reseller forwards each request through its own account pool (or another upstream supplier) and returns the answer in the familiar API format.

Because the response looks exactly like the official API, most developers never notice a third party sitting in the middle, able to read every prompt and every answer. According to Startup Fortune, this supply chain is split into specialised links and advertised openly on GitHub, Taobao and Telegram.

Why is it so cheap?

There’s no pricing magic. Research by Zilan Qian (Oxford China Policy Lab), reported by Tom’s Hardware and Let’s Data Science, identifies the usual sources of supply:

On top of that, many resellers profit from your data itself — and that is the part to worry about most.

The 5 risks you are really buying

1. Prompts and outputs are logged and sold as training data

The reseller sees both sides of every request. Developer prompts routinely contain source code, database schemas, support tickets, contracts, product plans and sometimes customer records. According to the reports above, this data can be stored and resold as training data, or used for distillation — training a cheaper model on a stronger model’s answers.

2. The model gets swapped

You pay for Opus but may be served by a cheaper model, or have your context silently truncated. An arXiv paper from March 2026, Real Money, Fake Models, audited 17 “shadow API” services: 45.83% of model-fingerprint tests failed to match the advertised model, and performance diverged by up to 47.21%. This kind of failure is hard to debug because it appears unevenly — good one moment, bad the next.

3. Terms violations — the service can die any day

On 19 February 2026, Anthropic updated its documentation to clarify that OAuth tokens from the Free, Pro and Max plans are only for Claude Code and Claude.ai. Using them in other products, tools or services — including the Agent SDK — breaches the consumer terms. Anthropic also does not allow third parties to offer Claude.ai logins or route other people’s requests through subscription credentials. Anthropic later stressed that personal experimentation with the Agent SDK or claude -p is still encouraged, but businesses built on top of it must use an API key.

4. Compliance and customer-contract exposure

If your company has promised customers (or a security review, or an enterprise contract) that data only flows through approved infrastructure, quietly routing it through an unofficial reseller breaks that promise — even if nothing ever goes wrong.

5. No continuity

Prepaid balances, anonymous operators, no invoices, no SLA. When an upstream account is banned or the reseller disappears, you lose both your balance and your service — with nobody to complain to.

Checklist before using any intermediary

If you must use an AI intermediary, ask it to answer these five questions clearly. If it can’t, keep customer data away from it:

1. Who is your upstream (Anthropic directly, AWS Bedrock, Google Cloud, Microsoft Foundry)?
2. Do you log prompts/responses? For how long? Who can access them?
3. Is there a contract, a proper invoice and a DPA (data processing agreement)?
4. What are your security controls and incident process?
5. How can I verify the model answering is the model you advertise?

Cutting Claude costs the legitimate way

Batching and caching stack: an overnight job that both batches requests and caches shared context gets considerably cheaper without trading away your data.

How to apply it

• Developers: Audit every place base_url is configured in code and environment variables (ANTHROPIC_BASE_URL). If you see an unfamiliar domain, find out where it came from before shipping.
• Tech Leads: Add “approved LLM providers” to your code-review and security-review checklists, just like an approved-library list.
• PMs: When budgeting for AI, keep two lines: tooling for the team (subscriptions) and product operations (the official API). Don’t let cost pressure push the team into the grey market.
• Personal experiments: If you still want to try a cheap service, use dummy data only — never company code or customer data.