The best mental model for Claude pricing is a utility bill. Your electricity company does not charge per appliance or per hour of use; it meters kilowatt-hours consumed. Claude works the same way, except the metered unit is the token, roughly a short word or word fragment of text, and there are two meters running: one counting the text you send (input tokens) and one counting the text the model writes back (output tokens). No subscriptions, no per-seat fees, no per-request charge. Usage times rate, twice, summed.
The two meters
Input tokens count everything you send with a request: your system prompt (the standing instructions), the user's message, any documents you include, and prior conversation turns you resend for context. This last part surprises people: the API is stateless, so a chat application resends the conversation history with each turn, and the input meter counts it every time. Long conversations get more expensive per turn as they grow.
Output tokens count what the model generates. Notice in the price list below that output tokens cost roughly five times more than input tokens. That asymmetry drives a practical rule: workloads that read a lot and write a little (summarize this 30-page contract into one paragraph) are dominated by input cost, while workloads that write a lot (draft a full report) are dominated by output cost. Knowing which kind you have tells you which knob matters.
The price list
These are the list prices per one million tokens as of July 2026. A useful feature of the 3P model: cloud marketplace list prices match Anthropic's first-party list prices, so you are not paying a platform markup by buying through your cloud. Committed-use discounts, where available, are negotiated with the cloud provider, not with Anthropic.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Notes |
|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | Most capable general model; default recommendation |
| Claude Sonnet 5 | $3.00 ($2.00 intro through Aug 31, 2026) | $15.00 ($10.00 intro) | Balanced workhorse |
| Claude Haiku 4.5 | $1.00 | $5.00 | Fast/cheap; 200K context (others 1M) |
To make the numbers concrete with a deliberately simple illustration: at Haiku 4.5 rates, a million tokens of input, which is on the order of a decent stack of novels, costs one dollar. Individual requests are typically fractions of a cent; bills grow because request counts grow.
Reading your own meter
You never have to guess token counts: every API response reports exactly what was metered. This is the same on all four platforms, since they share the same API shape and SDK:
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="my-project", region="global")
response = client.messages.create(
model="claude-sonnet-5",
max_tokens=300,
messages=[{"role": "user", "content": "Summarize this policy update: ..."}],
)
print(response.usage.input_tokens, response.usage.output_tokens)Logging these two numbers per request, tagged by feature and team, is the foundation of every cost practice worth having. Multiply by the rates above and you can reconstruct your bill line by line, forecast next month from request growth, and spot anomalies (a prompt that doubled in size, a feature stuck in a retry loop) days before the invoice would reveal them.
What moves the bill besides usage
Three levers change the effective rate without changing your workload. Model choice is the big one: the same request costs five times more on Opus 4.8 than on Haiku 4.5 per input token, so routing routine work to cheaper tiers matters. Prompt caching, available on all four platforms, lets you reuse long, repeated prompt sections (like a lengthy system prompt or shared document) at reduced cost instead of paying full input price on every request. Intro and negotiated pricing: Sonnet 5 carries introductory rates through August 31, 2026, and committed-use arrangements are a procurement conversation with your cloud provider. Budget against list prices and treat discounts as upside.
Where to go next
To turn these rates into a monthly forecast, work through Estimating Your Monthly Claude Bill Before You Commit. The caching lever gets a full article in Prompt Caching: The Single Biggest Cost Lever, and the unit itself is explained in Tokens 101.