Choosing a relay
How to choose a Claude Code relay
Relay taglines all sound alike. The differences are where you cannot see them: who the upstream is, how billing works, whether the cache hits. You can check each of the seven points below yourself, without taking anyone's word for it, ours included. Still deciding between a subscription, the API and a relay? Start with the three compared.
- You can check all seven points yourself: upstream, billing basis, cache hits, protocol pass-through, reliability record, top-ups and terms, and support.
- Verify the upstream with the checking script in a few minutes for a few cents.
- Cache reads cost 2.5–10% of the input price (5% on Opus 5.5 and Sonnet 5.5), so whether a session stays on one account often matters more to your bill than the unit price.
- A relay that asks you to set
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETASorENABLE_TOOL_SEARCH=falseis not passing Claude Code's requests through unchanged. - Start small and read the terms before deciding how much to top up.
1. Can the upstream be verified?
"Direct to the official API" is just a sentence. What can be verified is behavior: whether the prompt cache is written and read back, and whether thinking-block signatures pass Anthropic's verification. Run the checking script; it takes a few minutes and a few cents.
When a price looks too good to be true, check before you top up.
2. Does billing match Anthropic's?
Anthropic prices four kinds of tokens separately: input, output, cache writes and cache reads. Cache reads cost 2.5–10% of the input price (5% on Opus 5.5 and Sonnet 5.5), and most of Claude Code's context goes through them (multipliers in Anthropic's prompt caching pricing). Ask three questions about a relay's billing:
- Are the four kinds priced separately, at the same ratios as Anthropic's?
- Can you look up the usage and cost of every single request in the logs?
- What price are the "ratio" and "discount" relative to? Pricing one real session's bill at Anthropic's unit prices is more reliable than an advertised discount.
3. Does the cache hit reliably?
Claude Code resends the system prompt, tool definitions and files it has read on every turn. On a cache hit that part costs 5% of the input price on Opus 5.5 and Sonnet 5.5; on another account the cache has to be written again. Whether one session always lands on the same upstream account often matters more to what you pay than the unit price.
To check: work in Claude Code for twenty or thirty turns, then look at the share of input that came from cache reads in the usage logs. In a long session it should be high; if most of it is cache writes, requests are hopping between accounts. How to read this is in Caching and billing.
4. Is the protocol passed through unchanged?
Claude Code relies on Anthropic beta headers for the 1-hour cache, MCP tool search and more. Some relays cannot forward these headers, so they tell you to set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS or ENABLE_TOOL_SEARCH=false. Those settings stop the errors at the cost of fewer features and a cache that breaks more often. A relay that asks for them is not passing requests through unchanged.
5. Is reliability on public record?
"Stable" should come with data: a public status page with the success rate and latency over the last day and week, and whether requests in progress survive a release or restart. With only "all fine" in a group chat, you cannot tell whether a problem is your setup or their service.
6. Top-ups, refunds and terms
- Start small. Finish one real task before deciding, rather than prepaying a large amount up front.
- Read the terms of service: whether the balance expires, whether it is refundable, and what happens if the service ends.
- Know the payment methods and when credit arrives, and keep your order records.
7. Who helps when something breaks?
Is there a way to reach a person, and can they find the matching log from the time of your failed request? How complete the setup docs are also shows whether a relay is run with care.
How CoreCode does it
| Point | CoreCode |
|---|---|
| Upstream | Official Anthropic Claude accounts only. Verify it yourself with the checking script. |
| Billing | Input, output, cache writes and cache reads priced separately, on the same basis as Anthropic, with every request in the usage logs. Unit prices are on Pricing. |
| Cache | Each Claude Code session ID stays on one account, and the binding is restored after a restart. See Caching and billing. |
| Protocol | The native Messages API, with beta headers passed through unchanged and no compatibility settings needed. |
| Reliability | A public service status page with 24-hour and 7-day success rate and first-byte latency. Requests in progress survive releases. |
| Top-ups | ¥1 buys $1 of credit, from ¥10; the newcomer pack is ¥29.9 for $50, valid 7 days. See Top-up and the Terms of service. |
| Support | Errors, the FAQ and the WeChat group. |
FAQ
Is the cheapest relay the best one?
Not necessarily. A relay whose cache never hits can cost more in practice than one with a higher unit price and a cache that keeps hitting. Read price together with the cache hit rate; see does the cache hit reliably.
How do I tell whether a relay swaps models?
Use the checking script: it checks the response shape, that the cache is written and read back, and that a tampered thinking signature is rejected.
Why do some relays ask me to turn off beta features?
Because they cannot forward Anthropic's beta headers. Turning them off stops the errors, but the 1-hour cache, MCP tool search and more stop working, and the cache breaks more often.