Claude Code Max Plan vs a Pay-Per-Token API Proxy

Two different ways of paying

A flat plan gives you a fixed monthly price and a usage allowance that resets on a schedule. You never see a per-token bill, and you can stop thinking about cost until you hit the allowance. Check Anthropic's plan page for what the Max tiers cost and allow today, since both change.

A pay-per-token API works the opposite way. There is no allowance, no reset and no subscription. You top up credit, every request deducts a few fractions of a cent, and unused credit stays on the account. For this service, credit never expires, and failed or refused requests are not charged.

Neither is better in general. The right question is how many tokens you really use in a month. Most people have never counted, so the next section does it for three typical patterns.

What the proxy route actually involves

Claude Code talks Anthropic's API format. This service, at https://api.proxycodeclaude.com/v1, speaks the OpenAI format, so a local translation proxy sits between them. The usual tools are LiteLLM and router utilities for Claude Code; the configuration is a few lines, and the backend model ID is uncensored.

Be clear about what changes. The model behind this API is a single open-weight model from an independent provider, tuned to answer without content refusals for lawful adult use. It is not a Claude model. Your Max plan, if you have one, is a subscription to Anthropic's service, and it is not used while the CLI points at a different endpoint. Pointing the CLI elsewhere is a switch you control with environment variables, not an add-on.

Three usage profiles, priced exactly

Coding agents re-send the conversation on every turn, so input tokens dominate. The profiles below assume the whole input is billed at full price on each request, since no prompt caching discount is documented for this API.

Now the arithmetic, at $0.15 per 1M input tokens and $0.50 per 1M output tokens.

• Light. 20 x 30 = 600 requests. Input: 600 x 20,000 = 12,000,000 tokens, which is 12 x $0.15 = $1.80. Output: 600 x 600 = 360,000 tokens, which is 0.36 x $0.50 = $0.18. Total $1.98.
• Regular. 60 x 50 = 3,000 requests. Input: 3,000 x 30,000 = 90,000,000 tokens, or 90 x $0.15 = $13.50. Output: 3,000 x 800 = 2,400,000 tokens, or 2.4 x $0.50 = $1.20. Total $14.70.
• Heavy. 120 x 80 = 9,600 requests. Input: 9,600 x 45,000 = 432,000,000 tokens, or 432 x $0.15 = $64.80. Output: 9,600 x 1,000 = 9,600,000 tokens, or 9.6 x $0.50 = $4.80. Total $69.60.

Put your plan's monthly price next to those numbers. If it is well above the profile that looks like you, the pay-per-token route is cheaper on paper. If you are heavy and the plan is lower than $69.60, the flat plan wins on price alone. Your own request sizes will differ, so treat these as templates and replace the numbers.

What you give up

Price is only one column. Here is the rest.

• The model. A frontier Claude model and an open-weight model are not interchangeable on hard multi-file work. Tool calls are supported by this API, but how well the model drives an agent loop is for you to test, and no benchmark is claimed here.
• Anthropic-only features. Server-side tools, extended thinking controls and similar platform features have no equivalent in an OpenAI-format text API. Images are not supported either; the API is text only.
• Context. The window is 100,000 tokens for prompt plus completion. Use /compact before you get close.
• Parallelism. A full key allows 8 concurrent requests and 300 per minute. Trial keys allow 2 and 60.

Before you cancel or downgrade anything, run a week of real tasks through the alternative and compare the results, not only the invoice. A cheaper session that needs three retries is not cheaper.

What you gain is a model that does not stop at content refusals for lawful adult use, a bill that follows usage down to zero in a quiet month, and prompts that are not used for training. For a writer or researcher whose work keeps hitting refusals, that matters more than the price.

A hybrid setup that costs nothing extra

You do not have to pick. Claude Code reads its endpoint from environment variables, so two shell aliases give you both worlds:

Copy# normal Claude Code, using your existing plan or key
alias cc='claude'

# Claude Code through the local proxy to https://api.proxycodeclaude.com/v1
alias cc-alt='ANTHROPIC_BASE_URL=http://localhost:4000 ANTHROPIC_AUTH_TOKEN=sk-local-change-me ANTHROPIC_MODEL=uncensored claude'

Use cc for work where Claude's quality matters, and cc-alt for tasks that keep getting declined or for long, cheap, repetitive sessions. The proxy itself is a LiteLLM config with model: openai/uncensored and api_base: https://api.proxycodeclaude.com/v1; the documentation shows the request format it will send.

Getting started and budgeting

Create an account on the API key page and claim the $0.50 trial credit, valid for 7 days with no payment details. That covers about 5 sessions of the Light profile above, since one such session costs $1.98 / 20 = $0.099. When you are ready to continue, top ups run from $20 to $500 in whole amounts, paid in crypto only, with +5% from $50, +10% from $100 and +15% from $200. A $100 top-up, for example, becomes $110 of credit.

That $110 covers about 7 months of the Regular profile at $14.70 a month ($110 / $14.70 = 7.5). Rates and bonuses are listed on the pricing page, and Price claude vs Our Uncensored Proxy API goes deeper on per-token comparisons.

More guides

• Claude Code Proxy in Docker: Compose File and Settings
• Price claude vs Our Uncensored Proxy API