Claude Code on Windows: A PowerShell Cost Receipt

01 / Start with a receipt, not a monthly promise

A useful Claude Code budget has four lines: ordinary input, output, cache writes, and cache reads. Counting prompts alone tells you little about the final bill. Claude Code can report these token categories, but its displayed dollar estimate may use list prices rather than your relay’s actual rates. Treat the relay’s billing record as the receipt. ([code.claude.com](https://code.claude.com/docs/en/costs))

CheapRelay offers pay-as-you-go access with no monthly fee or subscription, billing real tokens, including cache reads and writes. That removes a recurring subscription charge from this setup; it does not make individual coding tasks free or predictably priced.

02 / Set the PowerShell environment before launching Claude

Assume Claude Code is already installed. In PowerShell, environment assignments use $env:NAME, not Bash’s export syntax. Set them in the same terminal that will launch Claude Code: child processes inherit that terminal’s environment. ([learn.microsoft.com](https://learn.microsoft.com/en-us/powershell/module/microsoft.powershell.core/about/about_environment_variables?view=powershell-7.1&utm_source=openai))

Run these commands individually. The prompts avoid putting a literal credential into the command itself; use a private terminal when entering it.

• ■ $env:ANTHROPIC_BASE_URL = Read-Host 'Paste the CheapRelay Claude Code API base'
• ■ $env:ANTHROPIC_AUTH_TOKEN = Read-Host 'Paste your CheapRelay API key'
• ■ $env:ANTHROPIC_MODEL = Read-Host 'Paste the exact supported model ID'
• ■ claude

Copy the API base and model identifier from CheapRelay’s current Claude Code instructions. Do not substitute a display name or assume an OpenAI-compatible endpoint is automatically the correct Claude Code endpoint. Anthropic documents ANTHROPIC_BASE_URL for gateway routing and ANTHROPIC_AUTH_TOKEN for gateway authentication; the gateway must support the required API format. ([code.claude.com](https://code.claude.com/docs/en/llm-gateway-connect))

For example, if choosing Claude Haiku 4.5, copy its exact relay model ID rather than guessing the spelling. Keep this first session temporary until a small request succeeds.

03 / Keep the endpoint persistent without saving the key

For the current process, check the endpoint with $env:ANTHROPIC_BASE_URL. Check credential presence without printing it using [bool]$env:ANTHROPIC_AUTH_TOKEN. These checks establish what the shell contains, not whether the server accepts the credential.

To persist only the endpoint, run [Environment]::SetEnvironmentVariable('ANTHROPIC_BASE_URL', $env:ANTHROPIC_BASE_URL, 'User'). Windows user-scoped variables survive the current process; $env: assignments alone do not. Restart the terminal application before testing inheritance in a fresh session. ([learn.microsoft.com](https://learn.microsoft.com/en-us/powershell/module/microsoft.powershell.core/about/about_Environment_Variables?utm_source=openai))

I recommend entering the key again for each working session rather than adding it to a repository or profile. If authentication fails, inspect the configured endpoint and credential method before sending repeated test requests.

04 / A four-line example receipt totaling $0.235

The following rates are hypothetical arithmetic inputs, not verified CheapRelay prices, advertised discounts, or measured task costs. Replace every rate with the current rate for your selected model and cache duration before budgeting.

• ■ Ordinary input: 40,000 tokens at an assumed $1.00 per million = $0.040.
• ■ Output: 10,000 tokens at an assumed $5.00 per million = $0.050.
• ■ Cache writes: 100,000 tokens at an assumed $1.25 per million = $0.125.
• ■ Cache reads: 200,000 tokens at an assumed $0.10 per million = $0.020.
• ■ Example usage total: $0.235, or 23.5 cents, before any applicable taxes or separate charges.

Each line follows the same calculation: tokens divided by 1,000,000, multiplied by the category’s rate. These buckets must be non-overlapping: do not charge cache tokens again as ordinary input. Caching has distinct read and write pricing, so one blended input rate can conceal the expensive part. ([claude.com](https://claude.com/resources/articles/maximizing-the-value-of-your-claude-code-sessions?c=42&fcdaa149_sort_Plus+ancien=asc&utm_source=openai))

05 / Turn one receipt into a working budget

Twenty identical example sessions would total $4.70; forty would total $9.40. Those are multiplication exercises, not monthly forecasts. Your next debugging task might consume a different amount of context or output.

Record the selected model, four token counts, applicable rates, and actual debit after your first small task. Compare the debit with your reconstruction before running unattended work. Sign up for CheapRelay when you are ready to test that first receipt, then budget from observed charges rather than a promised cost per prompt.