ClaudeAPI uses pay-as-you-go token billing in USD, with public Claude model rates currently set at 80% of Anthropic’s published list prices. There is no monthly subscription or minimum spend. Your balance is deducted according to the model, input tokens, output tokens, and any cache usage recorded for each request.
ClaudeAPI is an independent AI API gateway. It is not an Anthropic product or an authorized Anthropic reseller.
ClaudeAPI pricing comparison
Prices are shown per one million tokens. Input and output tokens are charged separately.
| Model | ClaudeAPI input | ClaudeAPI output | Anthropic list input | Anthropic list output |
|---|---|---|---|---|
| Claude Haiku 4.5 | $0.80 | $4 | $1 | $5 |
| Claude Sonnet 4.6 | $2.40 | $12 | $3 | $15 |
| Claude Opus 4.6 | $4 | $20 | $5 | $25 |
| Claude Opus 4.7 | $4 | $20 | $5 | $25 |
| Claude Opus 4.8 | $4 | $20 | $5 | $25 |
These are the rates displayed publicly by ClaudeAPI as of June 8, 2026. Check the console before deploying a high-volume workload because model availability and prices may change.
How ClaudeAPI billing works
ClaudeAPI calculates the cost of each request from its recorded token usage:
request cost =
(input tokens / 1,000,000 × input price)
+ (output tokens / 1,000,000 × output price)
+ applicable cache or tool charges
The API response includes a usage object containing the measured input and output token counts. You can also review usage and account costs in the ClaudeAPI console.
Example: a typical Sonnet request
Assume a Claude Sonnet 4.6 request uses:
200 input tokens
500 output tokens
No prompt caching or paid tools
The calculation is:
Input: 200 / 1,000,000 × $2.40 = $0.00048
Output: 500 / 1,000,000 × $12.00 = $0.00600
Total: $0.00648
Output tokens represent most of the cost in this example. Setting reasonable output limits can therefore have a larger effect than reducing a small prompt.
Choosing a Claude model by cost
The most expensive model is not automatically the best option for every request.
Workload Suggested model Why
Classification, extraction, routing, and short answers Haiku 4.5 Lowest token price
Coding, content generation, support, and general reasoning Sonnet 4.6 Balanced cost and capability
Complex agents, difficult reasoning, and demanding analysis Opus 4.8 Highest-capability tier
Mixed production traffic Haiku, Sonnet, and Opus Route each request by complexity
A practical routing strategy uses Haiku for lightweight preprocessing, Sonnet as the default model, and Opus only when the task requires additional reasoning depth.
Prompt caching prices
Prompt caching can reduce the cost of repeatedly sending the same prompt prefix, system instructions, documents, or conversation history.
The ClaudeAPI public pricing card currently lists the following cache-related rates for Claude Opus 4.8:
Cache operation Price per million tokens
Cache write $5
Cache read $0.50
Cache reads are substantially cheaper than normal Opus input tokens. Caching is most useful when a stable block of content is reused across multiple requests.
Typical examples include:
Long system prompts
Repeated reference documents
Shared few-shot examples
Stable agent instructions
Reused conversation prefixes
Cache duration and eligibility depend on the model and request format. Review the live console pricing before estimating a production caching workload.
Estimating common request costs
The following examples use 200 input tokens and 500 output tokens per request.
Model Input cost Output cost Approximate total
Haiku 4.5 $0.00016 $0.00200 $0.00216
Sonnet 4.6 $0.00048 $0.00600 $0.00648
Opus 4.8 $0.00080 $0.01000 $0.01080
These examples are calculations, not fixed prices per conversation. Actual costs depend on prompt length, generated output, cached tokens, tool usage, and the selected model.
How to reduce Claude API costs
Route requests by complexity
Use Haiku for predictable lightweight tasks, Sonnet for most application traffic, and Opus for requests that genuinely require deeper reasoning.
Limit unnecessary output
Output tokens cost more than input tokens across the main Claude model tiers. Use max_tokens, structured responses, and explicit length requirements to prevent unnecessarily long generations.
For example:
Return valid JSON with these fields only:
- category
- confidence
- explanation
Keep the explanation below 40 words.
Cache repeated content
Use prompt caching when the same instructions or reference material appear across many requests. A cache read can cost much less than processing the same input again.
Remove irrelevant context
Do not send an entire conversation or document when the current request only needs a small section. Summarizing old context can reduce both cost and latency.
Monitor actual usage
Estimate costs during development, but use the response usage data and console history for production decisions. Real traffic distributions are usually more useful than a single average-request estimate.
Payment and account balance
ClaudeAPI uses prepaid, pay-as-you-go account balances. Payments are processed through Stripe with support for major credit cards.
The public billing rules currently state:
No monthly subscription
No minimum usage commitment
Charges are based on actual API consumption
Account credits remain available until used
Taxes or payment-processing adjustments may apply where required
Enterprise billing is available for eligible teams
Do not calculate ClaudeAPI costs through an RMB multiplier or a foreign-exchange comparison. The international site displays and settles public pricing in USD.
Using ClaudeAPI with the Anthropic SDK
You can configure an Anthropic-compatible client by setting the ClaudeAPI gateway as the base_url and using your ClaudeAPI key.
import os
import anthropic
client = anthropic.Anthropic(
api_key=os.environ["CLAUDEAPI_KEY"],
base_url="https://gw.claudeapi.com",
)
message = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=512,
messages=[
{
"role": "user",
"content": "Explain token-based API billing in three bullet points.",
}
],
)
print(message.content)
print(message.usage)
Always use a model identifier currently available in your account. Supported models can change independently of an SDK release.
FAQ
Does ClaudeAPI charge a monthly fee?
No. ClaudeAPI currently uses pay-as-you-go billing without a monthly subscription or minimum spend.
Do ClaudeAPI credits expire?
The current public terms state that prepaid credits remain valid until they are consumed. Credits are generally non-refundable unless applicable law requires otherwise.
Why are input and output tokens priced differently?
Generated output requires more inference work, so output tokens have a higher per-token price across the main Claude model tiers.
Does extended thinking have a separate price?
Thinking tokens are generally counted as output tokens and billed at the model's output rate. A larger reasoning budget can therefore increase the total request cost.
Does prompt caching remove token charges?
No. Cache writes and cache reads have their own token rates. Cache reads are cheaper than processing normal input, but they are not free.
Which Claude model should I start with?
Sonnet 4.6 is a practical default for general coding, content, and application workloads. Use Haiku for lightweight high-volume tasks and Opus for requests requiring deeper reasoning.
Are enterprise prices the same as public prices?
Enterprise customers may have different billing arrangements, volume terms, or invoicing requirements. Confirm applicable rates with ClaudeAPI before estimating enterprise usage.
Summary
ClaudeAPI currently bills Claude requests in USD based on measured token usage. Public rates are displayed at 80% of Anthropic's published list prices: Haiku 4.5 costs $0.80/$4, Sonnet 4.6 costs $2.40/$12, and Opus 4.6 through 4.8 cost $4/$20 per million input/output tokens.
# Claude API pricing: ClaudeAPI billing rules for 2026
**ClaudeAPI uses pay-as-you-go token billing in USD, with public Claude model rates currently set at 80% of Anthropic's published list prices.** There is no monthly subscription or minimum spend. Your balance is deducted according to the selected model, input tokens, output tokens, and any cache usage recorded for each request.
[ClaudeAPI](https://claudeapi.com) is an independent AI API gateway. It is not an Anthropic product or an authorized Anthropic reseller.
## ClaudeAPI pricing comparison
Prices are shown per one million tokens. Input and output tokens are charged separately.
| Model | ClaudeAPI input | ClaudeAPI output | Anthropic list input | Anthropic list output |
| --- | ---: | ---: | ---: | ---: |
| Claude Haiku 4.5 | $0.80 | $4 | $1 | $5 |
| Claude Sonnet 4.6 | $2.40 | $12 | $3 | $15 |
| Claude Opus 4.6 | $4 | $20 | $5 | $25 |
| Claude Opus 4.7 | $4 | $20 | $5 | $25 |
| Claude Opus 4.8 | $4 | $20 | $5 | $25 |
These are the rates displayed by ClaudeAPI as of June 8, 2026. Check the console before deploying a high-volume workload because model availability and prices may change.
## How ClaudeAPI billing works
ClaudeAPI calculates the cost of each request from its recorded token usage:
```text
request cost =
(input tokens / 1,000,000 × input price)
+ (output tokens / 1,000,000 × output price)
+ applicable cache or tool charges
request cost =
(input tokens / 1,000,000 × input price)
+ (output tokens / 1,000,000 × output price)
+ applicable cache or tool charges
The API response includes a usage object containing the measured input and output token counts. You can also review usage and account costs in the ClaudeAPI console.
Example: a typical Sonnet request
Assume a Claude Sonnet 4.6 request uses:
200 input tokens
500 output tokens
No prompt caching or paid tools
The calculation is:
Input: 200 / 1,000,000 × $2.40 = $0.00048
Output: 500 / 1,000,000 × $12.00 = $0.00600
Total: $0.00648
Output tokens represent most of the cost in this example. Setting reasonable output limits can therefore have a larger effect than reducing a small prompt.
Choosing a Claude model by cost
The most expensive model is not automatically the best option for every request.
Workload Suggested model Why
Classification, extraction, routing, and short answers Haiku 4.5 Lowest token price
Coding, content generation, support, and general reasoning Sonnet 4.6 Balanced cost and capability
Complex agents, difficult reasoning, and demanding analysis Opus 4.8 Highest-capability tier
Mixed production traffic Haiku, Sonnet, and Opus Route each request by complexity
A practical routing strategy uses Haiku for lightweight preprocessing, Sonnet as the default model, and Opus only when the task requires additional reasoning depth.
Prompt caching prices
Prompt caching can reduce the cost of repeatedly sending the same prompt prefix, system instructions, documents, or conversation history.
The ClaudeAPI public pricing card currently lists the following cache-related rates for Claude Opus 4.8:
Cache operation Price per million tokens
Cache write $5
Cache read $0.50
Cache reads are substantially cheaper than normal Opus input tokens. Caching is most useful when a stable block of content is reused across multiple requests.
Typical examples include:
Long system prompts
Repeated reference documents
Shared few-shot examples
Stable agent instructions
Reused conversation prefixes
Cache duration and eligibility depend on the model and request format. Review the live console pricing before estimating a production caching workload.
Estimating common request costs
The following examples use 200 input tokens and 500 output tokens per request.
Model Input cost Output cost Approximate total
Haiku 4.5 $0.00016 $0.00200 $0.00216
Sonnet 4.6 $0.00048 $0.00600 $0.00648
Opus 4.8 $0.00080 $0.01000 $0.01080
These examples are calculations, not fixed prices per conversation. Actual costs depend on prompt length, generated output, cached tokens, tool usage, and the selected model.
How to reduce Claude API costs
Route requests by complexity
Use Haiku for predictable lightweight tasks, Sonnet for most application traffic, and Opus for requests that genuinely require deeper reasoning.
Limit unnecessary output
Output tokens cost more than input tokens across the main Claude model tiers. Use max_tokens, structured responses, and explicit length requirements to prevent unnecessarily long generations.
For example:
Return valid JSON with these fields only:
- category
- confidence
- explanation
Keep the explanation below 40 words.
Cache repeated content
Use prompt caching when the same instructions or reference material appear across many requests. A cache read can cost much less than processing the same input again.
Remove irrelevant context
Do not send an entire conversation or document when the current request only needs a small section. Summarizing old context can reduce both cost and latency.
Monitor actual usage
Estimate costs during development, but use the response usage data and console history for production decisions. Real traffic distributions are usually more useful than a single average-request estimate.
Payment and account balance
ClaudeAPI uses prepaid, pay-as-you-go account balances. Payments are processed through Stripe with support for major credit cards.
The public billing rules currently state:
No monthly subscription
No minimum usage commitment
Charges are based on actual API consumption
Account credits remain available until used
Taxes or payment-processing adjustments may apply where required
Enterprise billing is available for eligible teams
Do not calculate ClaudeAPI costs through an RMB multiplier or a foreign-exchange comparison. The international site displays and settles public pricing in USD.
Using ClaudeAPI with the Anthropic SDK
You can configure an Anthropic-compatible client by setting the ClaudeAPI gateway as the base_url and using your ClaudeAPI key.
import os
import anthropic
client = anthropic.Anthropic(
api_key=os.environ["CLAUDEAPI_KEY"],
base_url="https://gw.claudeapi.com",
)
message = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=512,
messages=[
{
"role": "user",
"content": "Explain token-based API billing in three bullet points.",
}
],
)
print(message.content)
print(message.usage)
Always use a model identifier currently available in your account. Supported models can change independently of an SDK release.
FAQ
Does ClaudeAPI charge a monthly fee?
No. ClaudeAPI currently uses pay-as-you-go billing without a monthly subscription or minimum spend.
Do ClaudeAPI credits expire?
The current public terms state that prepaid credits remain valid until they are consumed. Credits are generally non-refundable unless applicable law requires otherwise.
Why are input and output tokens priced differently?
Generated output requires more inference work, so output tokens have a higher per-token price across the main Claude model tiers.
Does extended thinking have a separate price?
Thinking tokens are generally counted as output tokens and billed at the model's output rate. A larger reasoning budget can therefore increase the total request cost.
Does prompt caching remove token charges?
No. Cache writes and cache reads have their own token rates. Cache reads are cheaper than processing normal input, but they are not free.
Which Claude model should I start with?
Sonnet 4.6 is a practical default for general coding, content, and application workloads. Use Haiku for lightweight high-volume tasks and Opus for requests requiring deeper reasoning.
Are enterprise prices the same as public prices?
Enterprise customers may have different billing arrangements, volume terms, or invoicing requirements. Confirm applicable rates with ClaudeAPI before estimating enterprise usage.
Summary
ClaudeAPI currently bills Claude requests in USD based on measured token usage. Public rates are displayed at 80% of Anthropic's published list prices: Haiku 4.5 costs $0.80/$4, Sonnet 4.6 costs $2.40/$12, and Opus 4.6 through 4.8 cost $4/$20 per million input/output tokens.
# Claude API pricing: ClaudeAPI billing rules for 2026
**ClaudeAPI uses pay-as-you-go token billing in USD, with public Claude model rates currently set at 80% of Anthropic's published list prices.** There is no monthly subscription or minimum spend. Your balance is deducted according to the selected model, input tokens, output tokens, and any cache usage recorded for each request.
[ClaudeAPI](https://claudeapi.com) is an independent AI API gateway. It is not an Anthropic product or an authorized Anthropic reseller.
## ClaudeAPI pricing comparison
Prices are shown per one million tokens. Input and output tokens are charged separately.
| Model | ClaudeAPI input | ClaudeAPI output | Anthropic list input | Anthropic list output |
| --- | ---: | ---: | ---: | ---: |
| Claude Haiku 4.5 | $0.80 | $4 | $1 | $5 |
| Claude Sonnet 4.6 | $2.40 | $12 | $3 | $15 |
| Claude Opus 4.6 | $4 | $20 | $5 | $25 |
| Claude Opus 4.7 | $4 | $20 | $5 | $25 |
| Claude Opus 4.8 | $4 | $20 | $5 | $25 |
These are the rates displayed by ClaudeAPI as of June 8, 2026. Check the console before deploying a high-volume workload because model availability and prices may change.
## How ClaudeAPI billing works
ClaudeAPI calculates the cost of each request from its recorded token usage:
```text
request cost =
(input tokens / 1,000,000 × input price)
+ (output tokens / 1,000,000 × output price)
+ applicable cache or tool charges
The API response includes a usage object containing the measured input and output token counts. You can also review usage and account costs in the ClaudeAPI console.
Example: a typical Sonnet request
Assume a Claude Sonnet 4.6 request uses:
- 200 input tokens
- 500 output tokens
- No prompt caching or paid tools
The calculation is:
Input: 200 / 1,000,000 × $2.40 = $0.00048
Output: 500 / 1,000,000 × $12.00 = $0.00600
Total: $0.00648
Input: 200 / 1,000,000 × $2.40 = $0.00048
Output: 500 / 1,000,000 × $12.00 = $0.00600
Total: $0.00648
Output tokens represent most of the cost in this example. Setting reasonable output limits can therefore have a larger effect than reducing a small prompt.
Choosing a Claude model by cost
The most expensive model is not automatically the best option for every request.
| Workload | Suggested model | Why |
|---|---|---|
| Classification, extraction, routing, and short answers | Haiku 4.5 | Lowest token price |
| Coding, content generation, support, and general reasoning | Sonnet 4.6 | Balanced cost and capability |
| Complex agents, difficult reasoning, and demanding analysis | Opus 4.8 | Highest-capability tier |
| Mixed production traffic | Haiku, Sonnet, and Opus | Route each request by complexity |
A practical routing strategy uses Haiku for lightweight preprocessing, Sonnet as the default model, and Opus only when the task requires additional reasoning depth.
Prompt caching prices
Prompt caching can reduce the cost of repeatedly sending the same prompt prefix, system instructions, documents, or conversation history.
The ClaudeAPI pricing page currently lists these cache-related rates for Claude Opus 4.8:
| Cache operation | Price per million tokens |
|---|---|
| Cache write | $5 |
| Cache read | $0.50 |
Cache reads are cheaper than normal Opus input tokens. Caching is most useful when a stable block of content is reused across multiple requests.
Typical examples include:
- Long system prompts
- Repeated reference documents
- Shared few-shot examples
- Stable agent instructions
- Reused conversation prefixes
Cache duration and eligibility depend on the model and request format. Review the live console pricing before estimating a production caching workload.
Estimating common request costs
The following examples use 200 input tokens and 500 output tokens per request.
| Model | Input cost | Output cost | Approximate total |
|---|---|---|---|
| Haiku 4.5 | $0.00016 | $0.00200 | $0.00216 |
| Sonnet 4.6 | $0.00048 | $0.00600 | $0.00648 |
| Opus 4.8 | $0.00080 | $0.01000 | $0.01080 |
These calculations are examples, not fixed prices per conversation. Actual costs depend on prompt length, generated output, cached tokens, tool usage, and the selected model.
How to reduce Claude API costs
Route requests by complexity
Use Haiku for predictable lightweight tasks, Sonnet for most application traffic, and Opus for requests that genuinely require deeper reasoning.
Limit unnecessary output
Output tokens cost more than input tokens across the main Claude model tiers. Use max_tokens, structured responses, and explicit length requirements to prevent unnecessarily long generations.
For example:
Return valid JSON with these fields only:
- category
- confidence
- explanation
Keep the explanation below 40 words.
Return valid JSON with these fields only:
- category
- confidence
- explanation
Keep the explanation below 40 words.
Cache repeated content
Use prompt caching when the same instructions or reference material appear across many requests. Cache reads can cost much less than processing the same input again.
Remove irrelevant context
Do not send an entire conversation or document when the current request only needs a small section. Summarizing old context can reduce both cost and latency.
Monitor actual usage
Estimate costs during development, but use response usage data and console history for production decisions. Real traffic distributions are usually more useful than a single average-request estimate.
Payment and account balance
ClaudeAPI uses prepaid, pay-as-you-go account balances. Payments are processed through Stripe with support for major credit cards.
The current billing rules include:
- No monthly subscription
- No minimum usage commitment
- Charges based on actual API consumption
- Account credits remain available until used
- Taxes or payment-processing adjustments may apply where required
- Enterprise billing is available for eligible teams
Do not calculate ClaudeAPI costs through an RMB multiplier or foreign-exchange comparison. The international site displays and settles public pricing in USD.
Using ClaudeAPI with the Anthropic SDK
You can configure an Anthropic-compatible client by setting the ClaudeAPI gateway as the base_url and using your ClaudeAPI key.
import os
import anthropic
client = anthropic.Anthropic(
api_key=os.environ["CLAUDEAPI_KEY"],
base_url="https://gw.claudeapi.com",
)
message = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=512,
messages=[
{
"role": "user",
"content": "Explain token-based API billing in three bullet points.",
}
],
)
print(message.content)
print(message.usage)
import os
import anthropic
client = anthropic.Anthropic(
api_key=os.environ["CLAUDEAPI_KEY"],
base_url="https://gw.claudeapi.com",
)
message = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=512,
messages=[
{
"role": "user",
"content": "Explain token-based API billing in three bullet points.",
}
],
)
print(message.content)
print(message.usage)
Always use a model identifier currently available in your account. Supported models can change independently of an SDK release.
FAQ
Does ClaudeAPI charge a monthly fee?
No. ClaudeAPI currently uses pay-as-you-go billing without a monthly subscription or minimum spend.
Do ClaudeAPI credits expire?
The current public terms state that prepaid credits remain valid until they are consumed. Credits are generally non-refundable unless applicable law requires otherwise.
Why are input and output tokens priced differently?
Generated output requires more inference work, so output tokens have a higher per-token price across the main Claude model tiers.
Does extended thinking have a separate price?
Thinking tokens are generally counted as output tokens and billed at the model’s output rate. A larger reasoning budget can therefore increase the total request cost.
Does prompt caching remove token charges?
No. Cache writes and cache reads have their own token rates. Cache reads are cheaper than processing normal input, but they are not free.
Which Claude model should I start with?
Sonnet 4.6 is a practical default for general coding, content, and application workloads. Use Haiku for lightweight high-volume tasks and Opus for requests requiring deeper reasoning.
Are enterprise prices the same as public prices?
Enterprise customers may have different billing arrangements, volume terms, or invoicing requirements. Confirm applicable rates with ClaudeAPI before estimating enterprise usage.
Summary
ClaudeAPI currently bills Claude requests in USD based on measured token usage. Public rates are displayed at 80% of Anthropic’s published list prices: Haiku 4.5 costs $0.80/$4, Sonnet 4.6 costs $2.40/$12, and Opus 4.6 through 4.8 cost $4/$20 per million input/output tokens.
For accurate budgeting, calculate input, output, caching, and tool usage separately. Then compare the estimate with actual usage data from your application and the cost history shown in the console.
Sources
Get started
ClaudeAPI bills by usage with no monthly minimum or subscription. Get an API key and review the live model prices in the console before deploying.
