Home → Guides → Claude Opus 5 in production

Claude Opus 5 in production

This model's bill is rarely about the tokens you think it is about. Here is one of ours, taken apart line by line.

The short version

Cache writes cost more than fresh input, not less. Anthropic bills a cache read at 10% of the input rate and a cache write at 125%. Same tokens, opposite directions, twelve and a half times apart.
In one real session of ours, 81% of the bill was cache writes — context being rebuilt, producing no output at all. Full arithmetic below.
Before dropping a tier to cut costs, look at your cache-write share. If it is large, stabilising context saves more than switching models, and it does not require re-validating a single output.
List price is $5.00 per 1M input, $25.00 per 1M output.

A real invoice, taken apart

7 calls on this platform. Look at the token counts before the money:

input          90 tokens  ×  $5.00 / 1M            =  $0.00045
output      8,338 tokens  ×  $25.00 / 1M           =  $0.20845
cache read  7,082 tokens  ×  $0.50 / 1M   (10%)    =  $0.00354
cache write 147,302 tokens  ×  $6.25 / 1M  (125%)   =  $0.92
                                                      ─────────
                                                      $1.13308

invoice total for those 7 calls:                    $1.13

Ninety input tokens. Eight thousand output tokens. And 81% of the money went to cache writes — rewriting context that produced nothing a user ever saw.

That arithmetic also confirms Anthropic's cache multipliers against a real bill rather than a docs page: reads at a tenth of input, writes at one and a quarter times it.

Why this happens

Prompt caching works by storing a prefix of your prompt. Send the same prefix again and it is a cheap read. Change anything inside it — even one character near the top — and the whole prefix is written again at 125% of the input rate.

Things that quietly invalidate a cache prefix:

A timestamp, request ID or random seed near the start of the system prompt.
Tool definitions that get reordered between calls (many frameworks iterate a dict).
A file list or directory listing regenerated in a different order.
Conversation history trimmed from the front instead of the back.

Each of those turns a $0.50 read into a $6.25 write. In an agent loop that runs for hours, that difference is the entire bill.

The order of operations for cutting a Claude bill

Measure the cache-write share first. Your usage log has the numbers. If cache writes are a large fraction, stop here — the rest of the list is smaller.
Stabilise the prefix. Move anything variable to the end of the prompt. Sort tool definitions deterministically. Trim conversation history from the tail.
Shorten output. Output bills at five times input on this model. A prompt that produces 30% shorter answers saves more than most model swaps.
Only then consider dropping to claude-sonnet-5, which is $2.00 / $10.00 against $5.00 / $25.00 here — and which requires re-validating your outputs, unlike the three steps above.

When to use it, and when not

Reach for Opus when a task has genuinely defeated the cheaper models: multi-file refactors, long-context reasoning, coding agents working over a real codebase.

Do not send it classification, extraction, routing or tagging. Those are high-volume and low-judgement, they dominate token counts in most pipelines, and a small model handles them indistinguishably. Routing only the final reasoning step to Opus usually cuts blended cost by more than half.

Calling it

from openai import OpenAI

client = OpenAI(api_key="YOUR_KEY", base_url="https://apiclan.us/v1")

resp = client.chat.completions.create(
    model="claude-opus-5",
    messages=[{"role": "user", "content": "Hello"}],
)

Using Claude Code or the Anthropic SDK? Those clients append /v1/messages themselves, so the base URL is https://apiclan.us with no /v1. Getting it backwards returns a 404 from the edge — the request never reaches the gateway and never appears in your usage log, which looks exactly like a bad key. Per-client settings are in the base URL guide.

What you pay here

Per 1M tokens	List price	APICLAN	You save
Input	$5.00	$1.50	70%
Output	$25.00	$7.50	70%

Cache reads and writes bill at the same fraction of their own rates. Calculator and sibling models on the price page.

Questions people actually ask

My token count is small but the bill is not. Why?

Almost always cache writes. They bill above the input rate and do not appear in the input-token column. Check the cache-creation figure in your usage log.

Should I just turn caching off?

No. Reads are a tenth of the input rate — caching is heavily net-positive when the prefix is stable. The problem is never caching itself, it is a prefix that keeps changing.

Opus 5 or Sonnet 5?

Sonnet is $2.00 / $10.00 against Opus at $5.00 / $25.00 and is fast enough for interactive use. Try Sonnet first on twenty real inputs from your own product; move up only where you can see the difference.

Start using it

No subscription, no monthly minimum, no sales call. Top up with USDT and spend what you use — 1 USDT gives you 2 credits of API balance.

Read the 30-second quickstart

Prices quoted on this page are regenerated automatically from live billing data. Third-party terms are quoted from that party's own published documentation.