I Spent 28.67 CNY on 4.07 Billion Tokens: A Real Cost Breakdown for LLM API Relays

1261398983 — Wed, 30 Sep 2026 18:12:21 +0000

Cost claims about LLM API relays are usually unfalsifiable marketing. Here is a breakdown I can actually defend, because it comes from my own billing dashboard and from request-level data I pulled myself.

The headline numbers

Metric	Value
Requests	30,734
Tokens	4.07 billion
Paid	28.67 CNY (~4 USD)
List price equivalent	359.51 CNY (~50 USD)
Effective ratio	7.97%

That is a 92% reduction against list price. The interesting part is not the number, it is why it holds up.

Mechanism 1: prompt cache hits

This is the single most under-discussed lever. Here is a real call from my logs:

Metric	Tokens
prompt_tokens	1,755
prompt_cache_hit_tokens	1,536
prompt_cache_miss_tokens	219
completion_tokens	20

87.5% of the input was served from cache. For agentic workflows - Claude Code, Codex CLI, any tool that re-sends a large system prompt and a long file context on every turn - this dominates the bill. If your provider does not expose cache hit/miss split in the response, you are flying blind.

You can verify it yourself. The usage block in the response looks like this:

{
  "usage": {
    "prompt_tokens": 1755,
    "prompt_cache_hit_tokens": 1536,
    "prompt_cache_miss_tokens": 219,
    "completion_tokens": 20
  }
}

If those two cache fields are absent, ask your provider why.

Mechanism 2: group rate multipliers

Many relays expose tiered rate multipliers per model group. In my case the open-weight group bills at 0.08x of list price, and a separate premium group at 0.22x. The multiplier is published in the console, not hidden in a PDF.

Cache hits and multipliers compound: 0.08x applied on top of cache-priced input tokens. That is how you land at 8% of list.

What I actually run through one endpoint

deepseek-v4-flash      deepseek-v4.1-flash    deepseek-v4-pro
glm-5.2                glm-5.3                glm-5.3-flash
kimi-k2.8              kimi-k3                minimax-m3
hy3                    hy4

Both protocols terminate on the same base URL, which matters because it means one credential and one bill:

Endpoint	Protocol	Status	Latency
GET /v1/models	OpenAI	200	1.61s
POST /v1/chat/completions	OpenAI	200	2.19s
POST /v1/responses	OpenAI Responses	200	2.37s
POST /v1/messages	Anthropic	200	3.94s

How to sanity-check any relay before you commit

Send a request without an API key and confirm you get 401, not a helpful error page.
Call /v1/models and check the model list matches what the pricing page claims.
Read the usage block and confirm cache hit/miss fields exist.
Send the same prompt twice and verify the second call is cheaper. That is cache working.
Cross-check one model's answer against the official API if you can. This is the only real test of whether you are getting the model you paid for.

Caveats worth stating plainly

Aggregators are not the official vendor. No enterprise SLA, no data-residency guarantee. If you need those, pay official prices.
Cheap relays can and do disappear. Top up small amounts. I keep a month of runway at most.
Latency figures above come from one residential connection in Asia. Your mileage will differ.
I am on a referral program, so the last link pays me roughly 10%. The plain domain behaves identically.

Running Claude Code and Codex CLI Behind a Single OpenAI-Compatible Endpoint

1261398983 — Wed, 30 Sep 2026 17:19:28 +0000

If you are using AI coding agents like Claude Code or Codex CLI and you keep hitting the same three walls - regional access, overseas card requirements, and one API key per model - this post is a practical writeup of how I consolidated everything into a single endpoint.

The problem

My setup used to look like this:

Claude Code -> Anthropic key (needs an overseas card)
Codex CLI -> OpenAI key (needs an overseas card)
Local tools (Cherry Studio / LobeChat / NextChat) -> each with its own key

Three wallets, three billing dashboards, three points of failure.

The approach: one base_url, both protocols

The cleanest fix is a gateway that speaks both the OpenAI and the Anthropic protocol on the same host, so every tool points at one base_url.

export OPENAI_BASE_URL="https://api.dshapi.icu/v1"
export OPENAI_API_KEY="sk-..."

For Claude Code:

export ANTHROPIC_BASE_URL="https://api.dshapi.icu"
export ANTHROPIC_AUTH_TOKEN="sk-..."
export ANTHROPIC_MODEL="claude-sonnet-4-5"

That is the whole config. No per-tool proxy, no key juggling.

Endpoint verification (I actually ran these)

Endpoint	Protocol	Result	Latency
GET /v1/models	OpenAI	200	1.61s
POST /v1/chat/completions	OpenAI	200	2.19s
POST /v1/responses	OpenAI Responses	200	2.37s
POST /v1/messages	Anthropic	200	3.94s

Requesting /v1/models without a key correctly returns 401, so auth is enforced.

Quick smoke test:

curl https://api.dshapi.icu/v1/models -H "Authorization: Bearer $OPENAI_API_KEY"

curl https://api.dshapi.icu/v1/chat/completions \

  -H "Authorization: Bearer $OPENAI_API_KEY" \

  -H "Content-Type: application/json" \

  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"hi"}]}'

Cost: where the savings actually come from

Cost numbers without a mechanism are just marketing. Here is the mechanism, taken from a real call:

Metric	Value
prompt_tokens	1,755
prompt_cache_hit_tokens	1,536 (87.5%)
prompt_cache_miss_tokens	219
completion_tokens	20

Two things compound:

Prompt cache hits - 87.5% of input tokens were served from cache, which bills far below normal input.
Group multiplier - the open-weights group bills at 0.08x of list price.

Aggregated over 30,734 requests and 4.07B tokens: 28.67 CNY paid against 359.51 CNY at list price, i.e. 7.97%.

Models available

deepseek-v4-flash      deepseek-v4.1-flash    deepseek-v4-pro

glm-5.2                glm-5.3                glm-5.3-flash

kimi-k2.8              kimi-k3                minimax-m3

hy3                    hy4

Registration friction

Item	Requirement
Email	Any (QQ mail works)
Overseas credit card	Not required
Payment	WeChat Pay / Alipay
Billing	Pay-as-you-go, balance does not expire

Caveats

This is an aggregator, not the official vendor. If you need enterprise SLAs or data-residency guarantees, use the official APIs.
I am on the referral program, so the last link pays me roughly 10%. The plain domain works identically if you would rather avoid that.
Latency is region dependent; these measurements come from a residential connection in Asia.

DEV Community: 1261398983