Claude Code Subscription Mechanics: Why the $200/mo Plan Pays for Inference, Not the CLI
The {{Claude Code}} CLI tool itself is free to download. The $200/month Claude Max subscription pays for a quota of inference billed against Anthropic's API endpoints at favorable internal rates — roughly equivalent to ~$5,000/month of pay-as-you-go API usage at retail prices.
A frequently confused detail in disputes over Anthropic's third-party policy: the Claude Code command-line tool is free. Anyone can download and run it. What the $200/month Claude Max subscription buys is access to a metered pool of model inference, billed against subscription quota rather than per-request via the public API. Because Anthropic enjoys internal compute discounts that retail API customers do not, the effective value of a Max subscription is large. Industry estimates put the equivalent retail API usage purchasable for $200/month at roughly $5,000 — a ~25x leverage ratio. This is what makes the subscription attractive to heavy users and what makes the question of where that inference can be consumed strategically important. Anthropic's compliance documentation forbids using OAuth tokens obtained via Claude Code, Claude.ai free, Pro, or Max in any other product, tool, or service — explicitly including the Agent SDK. The user-facing implication is that subscription quota is locked to first-party clients. Critics' counterargument: subscribers are paying for inference, not for the specific CLI binary, so restricting where that inference can be consumed is rent-seeking on the client surface rather than on the underlying compute. The technical mechanism — OAuth tokens scoped to specific client identifiers — is how the lock is enforced. Bypassing it (token extraction, custom clients impersonating the official one) is what Anthropic treats as a terms of service violation, and what OpenCode was forced to stop doing in April 2026.