Anthropic pricing
Claude Opus 5 pricing
Claude Opus 5 bills per million tokens, metered separately for input and output. The current rates are below, along with the parts of the bill a per-token estimate leaves out.
Claude Opus 5 rates in full
All figures are USD per million tokens, taken from Anthropic’s own pricing documentation. The API model identifier is claude-opus-5.
| Per million tokens | Standard |
|---|---|
| Input | $5 |
| Cached input read | $0.50 |
| Cache write, 5 minute | $6.25 |
| Cache write, 1 hour | $10 |
| Output | $25 |
| Batch input | $2.50 |
| Batch output | $12.50 |
Context window: 1M tokens, with up to 128K tokens of output per request.
What the rate table does not tell you
Output costs five times input, the same ratio Sonnet 5 uses, so a workload that leans on long responses scales the bill the same way regardless of which current Claude model does the work.
The full 1M token context window is included at standard pricing; a 900k token request bills at the same per-token rate as a 9k token request.
Fast mode, in research preview, charges $10 input and $50 output per million tokens across the full context window in exchange for significantly faster responses.
Reasoning tokens are billed as output tokens, so raising the effort level raises the output bill without changing a single prompt.
A worked monthly example
Take 500,000 calls in a month at 3,000 input tokens and 400 output tokens each. That is 1.5 billion input tokens against 200 million output tokens, so input is 88% of the token volume.
| Line | Tokens | Rate | Cost |
|---|---|---|---|
| Input | 1,500M | $5 | $7,500 |
| Output | 200M | $25 | $5,000 |
| Total | 1,700M | $12,500 |
Output is 12% of the tokens and 40% of the bill. That inversion is why prompt trimming so often fails to move the invoice, and it is the single most common reason a forecast built on Claude Opus 5 comes in under the real number.
How this page stays current
Most third-party pricing tables are typed in once and quietly rot. This one is generated from a checked figure with the date and source recorded, and it is cross-checked on every deploy against the live model catalog Spendline uses to price real API traffic. If the two disagree, the site fails to build rather than publishing a stale number.
Frequently asked questions
How much does the Claude Opus 5 API cost?
Claude Opus 5 is billed per million tokens, metered separately for input and output. As of September 16, 2026 the rate is $5 per million input tokens and $25 per million output tokens, with cached input read at $0.50 and batch input at $2.50.
Why is my Claude Opus 5 bill higher than the per-token estimate?
Most estimates multiply expected tokens by the input rate and stop there. Output is priced at $25 against $5 for input, so responses usually dominate the bill even when prompts look larger. Reasoning tokens bill as output. Retries and agent fan-out pay full input cost for calls that produced nothing. And a cache that gets written but never read is pure overhead rather than a saving.
What is the cheapest way to run Claude Opus 5?
Batch processing halves both input and output for work that does not need an immediate answer, taking input to $2.50 and output to $12.50. Caching a stable prefix takes repeat input to $0.50, 10% of the standard rate, and pays for itself once the cached content is read more often than it is written. The larger lever is usually routing: sending the requests that do not need this tier to a cheaper model, and capping spend per customer before the call goes out.
How do I track Claude Opus 5 spend per customer or per feature?
The provider invoice is one number for the whole organization, so per-customer cost has to be attributed at the call. That means tagging every request with the customer, team, or agent it belongs to and recording the resolved cost against that tag in a ledger you can close monthly. Spendline does this by sitting in front of the API, which also lets a budget be enforced before the request is forwarded rather than reported after.
Know what this costs per customer, not just per token
A rate card tells you what a token costs. It does not tell you which customer, team, or agent spent it, or stop the one that is running away. Spendline sits in front of the API so every call is attributed and checked against a budget before it is forwarded.
Run the 5 minute assessment