Full-strength Claude,
undiluted.
Every request is answered by an official Anthropic Claude account, never a substitute model. You pay 28% of the official price, each session stays on one account, and prompt caching keeps hitting. All of it can be checked.
Base URLhttps://corecode.live
$ python3 check-relay.py --base-url https://corecode.livehttps://corecode.live model=claude-sonnet-5-5PASS shape Anthropic Messages responsePASS cache wrote 17,755 tokens, read 17,755 backPASS signature tampered signature rejected (HTTP 400)Result: all checks look like Anthropic
Smart, tasteful and fast
Full-strength Claude Code, answered by the original models. Planning, coding and UI work all run at Claude's real level; pick Opus and Opus answers, never a downgrade or a swap.
Check it yourself →Official quality at 28% of the price
No subscription or plans; top up what you need and pay for what you use. Every model's prices sit beside Anthropic's, every charge is in your console, and nothing is hidden.
See the prices →The same money goes further
Enterprise-grade cache optimization keeps each session hitting its cache, so repeated context is billed at the low cache price. The same money lasts about 1.5× as long as at an 80% cache rate, and long tasks gain the most.
Why it lasts →Official Claude, checked three ways
Asking a model what it is proves nothing; one line of system prompt makes any model call itself Claude. We check behaviour that only Anthropic's API has and nobody else can fake. The check script, check-relay.py, is public, and anyone can run it against any relay.
- PASSResponse shape
Anthropic Messages format, ids starting with msg_, and usage with the cache write and read fields.
- PASSCache read-back
A 17,755-token prompt cache is written, and the next request reads all 17,755 back.
- PASSThinking signature
A thinking block sent back with one character of its signature changed is rejected by Anthropic with HTTP 400. No other model can produce a signature that passes.
- Every 3 daysModel fingerprint
Signatures prove the upstream is Anthropic; the open-source ModelTrace fingerprint checks the model every 3 days, with results and raw data published.
Why the same top-up goes further here
Claude Code resends the system prompt, tool definitions and files it has read on every turn, and in a long session that is most of the tokens. On a cache hit that part costs 5% of the input price (Opus 5.5, Sonnet 5.5). A relay that switches accounts mid-session loses the cache, and you pay full price for the same content again. CoreCode keeps each session on one official account, so the cache stays valid.
Example: a 20-turn session resending 50K tokens of context each turn. With hits, the first turn is a one-hour cache write (input price ×2) and the other 19 are cache reads (×0.05). Opus 5.5 and Sonnet 5.5 as the example; multipliers from Anthropic's official pricing.
on the context part of the bill. The longer the session, the more you save.
- Cache readinput ×0.05
- 5-minute cache writeinput ×1.25
- 1-hour cache writeinput ×2
Every cent is accounted for
The console's usage log records each request: the model, the input, output and cache tokens, and what it cost.
| Time | Input | Cache write | Cache read | Output | Cost |
|---|---|---|---|---|---|
| 14:02:11 | 1,204 | 48,310 | 0 | 812 | $0.4990 |
| 14:03:40 | 356 | 1,920 | 48,310 | 1,530 | $0.0974 |
| 14:05:02 | 212 | 2,610 | 50,230 | 2,046 | $0.1232 |
Three turns of one Claude Code session on claude-opus-5-5, at the pricing page's rates for the 「Claude 满血（仅限CC）」 group, in dollars of credit (¥1 = $1). The first turn writes the context to the cache; every later turn costs under a quarter of it.
- Published prices
Input, output, cache write and cache read prices for every model, listed beside Anthropic's on the pricing page.
- No subscription, no plans
Top up what you need with Alipay or WeChat Pay, at ¥1 = $1 of credit.
- Itemized
Every request's usage and cost is in the usage log, ready to reconcile.
Use the tools you already have
Claude Code needs one address changed. GPT, Gemini and Grok use the OpenAI-compatible API with the same token.
Claude Code
CodexWhat people ask first
How do I know an official Claude is answering?
Run the checking script: Anthropic's response format, a cache that reads back what was written, and a tampered thinking signature being rejected cannot be imitated by another model. CoreCode fingerprints the model every 3 days with ModelTrace and publishes the results on the same page.
How do I pay? Is there a monthly fee?
Top up with Alipay or WeChat Pay at ¥1 = $1 of credit. There is no subscription; each request is charged for what it used.
Do I need a proxy in mainland China?
No. corecode.live connects directly from mainland China; point Claude Code's ANTHROPIC_BASE_URL at it. See Using it from China.
Is it cheaper than a Claude subscription or the API?
It depends on how much you use it and whether you can get an account and payment Anthropic accepts. This post compares the prices and limits of all three.
Which models are available?
Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, plus GPT, Gemini and Grok groups. See Supported models.
Prices, versions and practice
Input, output and cache prices per model, and the bill for one Claude Code request.
VersionsClaude Code latest version and release notesThe current version, how to update, and what changed recently.
ChoosingA subscription, the API or a relay?Prices, limits and when each fits.
Add our support on WeChat to join the group
Stuck connecting, want to talk about using Claude Code, or have ideas for CoreCode? Come say so in the group.
Search the ID in WeChat or scan the QR code to add our support contact; once they accept, they will add you to the group.
Change one address and it's pure Claude.
Sign up, top up, create a token, then put this line into Claude Code.
ANTHROPIC_BASE_URL=https://corecode.live