Claude Opus 5.5 pricing depends on the route you use. Anthropic's Standard API token rates are not the monthly price of Claude Code, and neither is automatically the price charged by a managed workspace.
For the direct Standard API, the published base rates are $4 per million uncached input tokens and $20 per million output tokens. Check the official model specification and pricing table for the processing mode, cache category, tools, and provider terms that apply to your request. Rates were checked October 10, 2026.
First identify who sends the bill
A subscription label does not imply unlimited Opus requests. An API balance does not necessarily fund a subscription's extra-usage meter. If Claude Code is the client, use the billing-path guide to identify the active route before estimating anything.
Calculate a simple uncached example
Suppose an API task uses 200,000 uncached input tokens and 20,000 billable output tokens across all its requests. Assume Standard processing, no cache creation or reads, no paid tools, and no other charges.
- Input: 0.2 × $4 = $0.80.
- Output: 0.02 × $20 = $0.40.
- Illustrative total: $1.20.
This is arithmetic, not an observed task or a promise that a specific coding job costs $1.20. A tool loop may send the repository context repeatedly. Repairs, retries, and other billable output can change the total.
Use the provider's reported usage categories rather than the size of the first prompt. If a request reads cached context, account for it separately instead of charging all tokens at the uncached rate. The coding-agent cost calculator guide explains how to keep the whole task in view.
Budget for an accepted result
Consider two hypothetical routes through the same task. One produces a valid patch on its first attempt. Another has a smaller first bill but needs several repairs and human investigation. The first-request price alone cannot tell you which route was cheaper in useful work.
Before comparing models, define what “accepted” means: the reproduced defect is fixed, required checks pass, no unrelated behavior changed, and a reviewer understands the risks. Record all attempts against that rule.
The Opus 5.5 versus Sonnet 5.5 comparison owns the model-selection question. This page owns the billing calculation; choosing a cheaper rate and choosing the better tool for a task are related but different decisions.
Set controls before a long run
Use a task budget and stop when it is reached. Inspect the route's usage dashboard during a small trial, especially when parallel sessions or large contexts are involved. Confirm what happens when a limit is hit, and preserve work before restarting.
Do not change to another person's credentials or remove an organization's cap to keep the task running. A cap may be an intentional authorization boundary, not a technical error.
Check Agent.Space separately
Look at Agent.Space's current pricing and the model selector for the intended harness. An Anthropic release does not establish availability in Agent.Space, and the example above is not an Agent.Space quote.
If the supported route fits your project, create a workspace and try one bounded task with a recorded acceptance result and usage total. Compare the complete task—not just the advertised token rate—before increasing spend.