Claude Pro/Max Subscription vs API: Costs, Limits, How to Choose

Anthropic sells Claude two ways: a monthly subscription (Pro and Max tiers) and a pay-as-you-go API. Both can power Claude Code, and heavy users argue endlessly about which is cheaper. The honest answer is that they fail in opposite directions — subscriptions cap how much you can use, the API caps nothing but bills every token. This article lays out the real differences, which profile fits which option, and how a gateway lets you combine the two instead of picking one.

How the Subscription Works

A Claude Pro or Max subscription is a fixed monthly fee for access through Anthropic's apps and Claude Code. The defining property is the usage window: you get a quota that depletes as you work and resets on a schedule. Higher tiers buy larger windows, not unlimited use.

What this means in practice:

• Cost is perfectly predictable — the bill never surprises you
• Usage is not: hit the window mid-task and you wait for the reset
• Access is personal — one account, not something you can share with a team or wire into a backend service
• You cannot call the models from your own code; it is app and Claude Code access, not an API credential

How API Pay-As-You-Go Works

The API bills per token: input, output, and cache traffic each at the model's rate. There is no usage window to exhaust — the constraint is your budget and the provider's rate limits.

• Cost scales with usage in both directions: light months are cheap, heavy months are not
• Works everywhere: Claude Code, Cursor, Cline, your own services, CI jobs
• Keys can be issued per person, per tool, per environment — this is what makes team usage manageable
• A runaway agent session turns directly into money, so budgets and monitoring matter

Side by Side

Which Should You Pick?

The subscription fits you if:

• You use Claude interactively a few hours a day and rarely hit the window
• You want one flat number on the credit card statement
• Claude Code and the chat apps are the only surfaces you need

The API fits you if:

• You run multiple tools — Claude Code plus Cursor or Cline — or call models from code
• Your usage is spiky: intense weeks, quiet weeks, where a flat fee is waste half the time
• You are provisioning a team and need per-member keys, budgets, and usage analytics rather than a pile of personal accounts

The painful middle is the heavy Claude Code user: intense agentic sessions chew through subscription windows fastest, but the same sessions are exactly the workload that makes raw API bills spike. This is the profile gateways exist for.

Combining Both: the Claude MAX Group

Zivv runs a Claude MAX group built on a pool of Claude subscription accounts. Claude Code talks to it over the native Anthropic protocol — same wire format, no client hacks — while the pool absorbs the per-account usage windows, so long sessions do not hit the rate-limit wall a single subscription does. You still pay as you go from a normal balance, topped up at ¥1 = $1, at rates far below official per-token pricing.

The configuration is the standard two variables:

export ANTHROPIC_BASE_URL=https://zivv.pro
export ANTHROPIC_AUTH_TOKEN=sk-your-key-here
claude

And the same key reaches the pay-as-you-go API models — claude-opus-4-8 for hard problems, claude-sonnet-5 for daily work, plus GPT and Gemini families — through the OpenAI-compatible endpoint:

curl https://zivv.pro/v1/chat/completions \
  -H "Authorization: Bearer sk-your-key-here" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "messages": [{"role": "user", "content": "Reply OK"}]
  }'

So the practical split becomes: Claude Code sessions ride the MAX group for subscription-style economics without the window anxiety, while everything else — other clients, scripts, other model families — bills per token on the same balance. Current models and rates are listed in Model Hub; Windows users can follow the Claude Code Windows setup for the environment-variable details.

What About Teams?

Subscriptions do not distribute: five engineers means five personal accounts, five separate windows, and zero visibility into who used what. On the API side of a gateway, a team shares one balance while each member gets an independent key with its own budget, and usage is reported per member, per key, and per model. That reporting is usually the deciding factor for anyone who has tried to reconcile AI spend across a team — details on the Teams page.

FAQ

Is the subscription "unlimited"? No tier is unlimited. Every tier has a usage window; larger tiers have larger windows. If you never hit yours, the subscription is effectively unlimited for you — heavy agentic use hits it.

Which is cheaper for Claude Code? Light interactive use: usually the subscription. Sustained heavy use: a gateway's pooled MAX group or discounted per-token billing, because you neither idle a flat fee nor pay official API rates.

Can I switch between the MAX group and API models with one key? Yes. The group determines what a key can reach; Claude Code uses the Anthropic endpoint while other clients use the OpenAI-compatible one, all on one balance.

Do I lose Claude Code features going through a gateway? No — the MAX group speaks the native Anthropic protocol, so Claude Code runs unmodified. Setup details live in the Claude Code guide.

Wrap-Up

Pick the subscription for predictable, personal, moderate use. Pick the API for multi-tool, programmatic, or team use. If you are a heavy Claude Code user squeezed between the two, the combined route — MAX group for sessions, pay-as-you-go for everything else — is the configuration that stops the argument. Create an account and test it against a week of your real usage.