Home → Guides → Claude Sonnet 5 in production

Claude Sonnet 5 in production

Per-token price tells you very little about a Claude bill. Here is what 130 real billed calls look like.

The short version

Median and mean cost per call differ by roughly three times. Budgeting from an average will mislead you in both directions.
94% of all input arrived as cache reads across 130 calls — billed at a tenth of the input rate.
Cache writes still took 50% of the bill. They cost 125% of the input rate, so a small number of them outweighs a large number of reads.
List price is $2.00 per 1M input, $10.00 per 1M output.

Measured over 130 calls

	Tokens	Billed at	Cost at list
Input	240,399	$2.00 / 1M	$0.48
Output	75,474	$10.00 / 1M	$0.75
Cache read	3,879,282	$0.20 / 1M (10%)	$0.78
Cache write	792,025	$2.50 / 1M (125%)	$1.98

Total at list price $3.99 across 130 calls. Regenerated automatically from billing data — if the numbers move, this table moves.

The three-times spread, and why it matters

Cost per call is not evenly distributed. Most calls are cheap because their context was already cached; a minority are expensive because they rebuilt it. The mean sits well above the median as a result.

The practical consequence: do not size a budget by multiplying an average by your expected call count. Take the median for a steady-state estimate and the mean for a worst case, and expect the real figure to land between them depending on how stable your prompts are.

Cache reads carried the workload; cache writes carried the cost

3,879,282 tokens arrived as cache reads — 94% of all input. At $0.20 per 1M against $2.00 for fresh input, that alone is the difference between a workable bill and an unworkable one.

But 792,025 tokens were cache writes, at $2.50 per 1M, and they accounted for 50% of the total. Writes are 12.5 times a read. A workload that rebuilds its prefix a few times an hour will spend more on rebuilding than on answering.

Where Sonnet 5 is the right default

It is the balanced Claude: substantially cheaper than Opus and fast enough for interactive use. For most product work — support replies, content generation, structured extraction, code assistance that is not a multi-hour agent run — it is the model to try first.

Move up to claude-opus-5 only where you can see the difference on your own inputs. Move down to a small model for classification, routing and tagging, which dominate token counts and do not need judgement.

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-sonnet-5",
    messages=[{"role": "user", "content": "Hello"}],
)

Claude Code and the Anthropic SDK use a different base URL: https://apiclan.us with no /v1, because those clients append /v1/messages themselves. A wrong path returns a 404 at the edge that never reaches your usage log. See the base URL guide.

What you pay here

Per 1M tokens	List price	APICLAN	You save
Input	$2.00	$0.600	70%
Output	$10.00	$3.00	70%

Calculator and sibling models on the price page.

Questions people actually ask

How do I estimate a monthly bill?

Take your expected calls per month and multiply by both the median and the mean cost per call. The answer lands between them. Anything more precise than that is false precision until you have your own traffic to measure.

Two identical-looking calls cost different amounts. Is that a bug?

No. One hit the cache and one rebuilt it. Reads bill at 10% of input, writes at 125% — that is the whole difference.

Sonnet 5 or Opus 5?

Start with Sonnet. It is $2.00 / $10.00 against Opus at $5.00 / $25.00. Run both on twenty real inputs from your own product and read the outputs; that settles it better than any benchmark table.

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.