Anthropic API pricing is billed per 1 million tokens, with separate rates for input and output. Claude Haiku 4.5 is the entry-level model at $1 per 1M input tokens, while the mid-range Claude Sonnet 5 costs $2 per 1M input tokens. High-performance models like Claude Opus 5 are priced at $5 per 1M input tokens, though older versions like Opus 4.1 command a higher price of $15.
Costs are also influenced by prompt caching. This feature allows users to pay a "Write" fee to store context and a significantly lower "Read" fee for subsequent requests. For example, on Claude Sonnet 5, the standard input rate of $2 drops to $0.20 per 1M tokens when reading from the cache.
How do prompt caching rates impact the budget?
Prompt caching introduces a two-tier pricing structure for input tokens. When you first send a large block of text to be cached, you pay a "Write" rate, such as $1.25 for Haiku 4.5 or $2.50 for Sonnet 5. Once cached, "Read" requests for that same data are discounted to $0.10 and $0.20 per 1M tokens, respectively. This makes long-context applications significantly more predictable for finance teams.
What is the price difference between model versions?
Selecting the specific model version is a primary cost driver. While Claude Opus 5 costs $5 per 1M input tokens, the legacy Opus 4.1 costs three times as much at $15. Similarly, output costs for Opus 5 are $25 per 1M tokens, compared to $75 for Opus 4.1. Organizations should verify which version of Sonnet or Opus they are calling to avoid unnecessary overhead.
Measuring the value of Anthropic's API requires looking past the invoice to see how token spend converts into operational efficiency. A more expensive model like Opus 5 might save more employee time on complex reasoning than a cheaper model, resulting in a higher net return. Using NetLift, finance teams can correlate API costs against a baseline of manual task completion to quantify the true payback of their AI adoption.
Sources
All pricing on this page comes from official vendor pages: