Claude API pricing:
what every model costs
Per-million-token prices for every Claude model, how cache reads, cache writes and thinking tokens are billed, and how to work out a monthly bill.
Claude API pricing is pay-per-token: you are billed for what goes into the model and what comes out, in US dollars per million tokens, and nothing for the months you don't call it. That sounds simple until the first invoice, because a single request is billed in four parts, each at its own rate, and the rates differ a lot between Opus, Sonnet and Haiku. This page shows every Claude model's price at Half Off Dev, where the four parts come from, and how to turn them into a monthly estimate.
Every Claude model and its price
The table is read live from the same catalog the API charges from, so it never goes stale. Prices are per million tokens at Half Off Dev, which is half of Anthropic's list price for each model.
| Model | Context | Input | Output | Cache read |
|---|---|---|---|---|
| Claude Fable 5.1 claude-fable-5.1 | 1M | $5.00 | $25.00 | $0.13 |
| Claude Opus 5.5 claude-opus-5.5 | 1M | $2.00 | $10.00 | $0.10 |
| Claude Sonnet 5.5 claude-sonnet-5.5 | 1M | $1.00 | $5.00 | $0.05 |
| Claude Haiku 5.5 claude-haiku-5.5 | 1M | $0.05 | $0.25 | $0.005 |
| Claude Fable 5 claude-fable-5 | 1M | $5.00 | $25.00 | $0.50 |
| Claude Opus 5 claude-opus-5 | 1M | $2.50 | $12.50 | $0.25 |
| Claude Sonnet 5 claude-sonnet-5 | 1M | $1.00 | $5.00 | $0.10 |
| Claude Opus 4.8 claude-opus-4.8 | 1M | $2.50 | $12.50 | $0.25 |
| Claude Opus 4.7 claude-opus-4.7 | 1M | $2.50 | $12.50 | $0.25 |
| Claude Opus 4.6 claude-opus-4.6 | 1M | $2.50 | $12.50 | $0.25 |
| Claude Sonnet 4.6 claude-sonnet-4.6 | 1M | $1.50 | $7.50 | $0.15 |
| Claude Opus 4.5 claude-opus-4.5 | 200K | $2.50 | $12.50 | $0.25 |
| Claude Sonnet 4.5 claude-sonnet-4.5 | 1M | $1.50 | $7.50 | $0.15 |
| Claude Haiku 4.5 claude-haiku-4.5 | 200K | $0.50 | $2.50 | $0.05 |
USD per 1M tokens, live from the catalog.
Click a row for that model's page with its context window and cache-write price. The same data is at GET https://halfoff.dev/v1/models, which needs no key.
The four parts of a Claude bill
Every Claude response reports its usage in four fields. Each one is a different price:
| Usage field | What it counts | Compared with plain input |
|---|---|---|
input_tokens | Prompt tokens after the last cache breakpoint | The base price |
cache_creation_input_tokens | Prompt tokens written to the cache by this call | A little more |
cache_read_input_tokens | Prompt tokens read back from the cache | A small fraction |
output_tokens | The answer, extended thinking included | Several times more |
On claude-sonnet-5.5, a cache read is one twentieth of the input price and an output token costs five times an input token. That ratio is why the bill of an agent or a long chat is dominated by output and by how much of the history comes from the cache. The cache columns of the table above show the same ratios for every other model.
Worked example: one turn of a coding chat
Take one turn on claude-sonnet-5.5 where most of the history is cached:
| Token kind | Tokens | You pay per 1M | Cost |
|---|---|---|---|
| Uncached input | 3,000 | $1.00 | $0.003000 |
| Cache read | 50,000 | $0.05 | $0.002500 |
| Cache write | 1,500 | $1.25 | $0.001875 |
| Output | 1,200 | $5.00 | $0.006000 |
| Total | $0.013375 |
Without the cache, the same 54,500 input tokens would cost $0.054500 on their own, about seven times what the cached version pays for input. Output is the largest line of the bill even though it is the smallest number of tokens.
From one request to a monthly bill
- Count tokens per request in each of the four parts. The API tells you after the first call; before that, a prompt is roughly four characters per token in English, and
POST /v1/messages/count_tokenscounts exactly. - Multiply by the price of the model, per million.
- Multiply by requests per month.
The LLM API cost calculator does these three steps for every model at once, including a cache share, so you can compare the Claude models on your own numbers. How to estimate LLM API cost walks through three typical workloads by hand.
What changes the price you pay
Prices climb with model size, and the gap between the smallest and the largest Claude is a multiple, not a percentage. Opus vs Sonnet vs Haiku says which jobs justify which.
Repeated context is the cheapest tokens you can buy. Prompt caching covers where to put the breakpoints.
Extended thinking is billed as output, so a high effort setting can make the hidden reasoning bigger than the visible answer.
Set max_tokens to what the task needs, and ask for the format you will use, not an essay.
Every price above is already half of the official one (see official vs Half Off Dev prices). The cheapest way to use the Claude API puts that next to the other savings.
Frequently asked questions
How much does the Claude API cost per month?+
It depends on tokens, not on a plan. A hobby project that sends a few hundred short requests a day costs cents on a small model. A coding agent used all day, or an app answering thousands of long conversations, costs far more. Put your tokens per request and requests per day into the calculator for a number.
Is there a free tier for the Claude API?+
Half Off Dev has no free tier and no subscription: you top up a prepaid balance and pay only for the tokens you use. Your first top-up is matched up to $5.
Why is my Claude bill higher than I estimated?+
Usually output or thinking tokens. Output costs several times input, and extended thinking is billed as output even when you don't read it. Check output_tokens in the response usage, and lower the effort setting or max_tokens before changing models.
Is the half price the same for every Claude model?+
Yes. Every Claude model, and every kind of token (input, output, cache reads and cache writes), is billed at 50% of the official list price. The live table above shows the result for each model.
Does prompt caching still work at half price?+
Yes, and the two stack. Cache reads are a fraction of the input price on any Claude model; Half Off Dev also pins a conversation to the channel that holds its cache, so a long session keeps hitting it.
Ready to try it? Create an account, top up a few dollars and follow the quickstart: one base URL and one key are the only changes to your code. Compare other vendors on GPT API pricing and Gemini API pricing, or browse the full model list.