# Anthropic's Agent SDK Credits: The June 2026 Rules, Explained Anthropic now includes metered Agent SDK credits with Claude plans: $20 on Pro, $100 on Max 5x, $200 on Max 20x. What they permit, what stays prohibited, and the math. *Published 2026-06-12 ยท https://proxyllm.ai/blog/anthropic-agent-sdk-credits-explained* In June 2026, Anthropic added metered Agent SDK credits to its consumer Claude plans: $20 of monthly credit on Pro, $100 on Max 5x, $200 on Max 20x. The credits cover personal programmatic use, meaning your own `claude -p` runs, the Agent SDK in your own projects, and Claude Code GitHub Actions, metered at Anthropic's published API rates. They do not change Anthropic's prohibition on third-party services offering Claude login or routing requests through plan credentials on behalf of users. This page lays out the rules, the numbers, and the arithmetic of a metered allowance. ## The credit, by plan | Plan | Monthly price | Agent SDK credit | Interactive Claude Code | | -------------- | ------------- | ---------------- | --------------------------- | | Claude Pro | $20 | $20/mo | draws on plan usage windows | | Claude Max 5x | $100 | $100/mo | draws on plan usage windows | | Claude Max 20x | $200 | $200/mo | draws on plan usage windows | The credit equals the plan price at each tier and resets with the monthly cycle. Programmatic usage meters against it at Anthropic's published API rates (anthropic.com/pricing), token by token, like an API budget that comes bundled with the plan. Interactive Claude Code sessions are a separate system: they keep drawing on the plan's normal usage windows, not on the credit. ## What the credits permit Three uses, all personal: - `claude -p` from your own scripts, cron jobs, and pipelines. - The Agent SDK in projects you build and run for yourself. - Claude Code GitHub Actions in your own repositories. The common thread is you, on your own account, for your own use. Before June 2026, Anthropic's position was that subscription OAuth tokens could not be used in any product or service outside Anthropic's own applications, which made even personal `claude -p` automation a violation of the consumer terms. The credits carve that personal shape back in, with a meter attached. ## What stays prohibited Anthropic's Claude Code documentation states it "does not permit third-party developers to offer Claude.ai login or to route requests through Free, Pro, or Max plan credentials on behalf of their users" (code.claude.com/docs/en/legal-and-compliance). The credits sit inside that boundary, not across it. Concretely: a hosted service signed in with your Claude subscription, a gateway exposing your Max plan as an API endpoint for an app someone else operates on your behalf, or any tool offering "log in with Claude" to its own customers remains out, regardless of who holds the credit. The credits restore personal programmatic Claude; they do not reopen the door for hosted services on subscription credentials. ## How the rules got here | Date | What changed | | ------------- | ------------------------------------------------------------------------------------------------------------------------------ | | January 2026 | Server-side checks cut third-party tools (OpenCode, Cline, RooCode, Goose) off from subscription OAuth, without advance notice | | February 2026 | Anthropic clarified that subscription OAuth tokens in any other product, tool, or service violate the Consumer Terms | | June 2026 | Agent SDK credits added to Pro and Max plans, restoring metered personal programmatic use | The fuller history, including what the January enforcement looked like for the tools it hit, is in [why we don't support Claude Code](/blog/why-no-claude-code-support). ## The arithmetic of a metered allowance A credit is a budget, not a window: it meters automation at API rates until it runs out, so cost tracks volume up to the cap and then bills per token. Take a nightly `claude -p` job that meters $4 per run, 30 runs a month, $120 of metered usage: | Plan | Credit | Nights covered | Overflow at API rates | | -------------- | ------ | -------------------- | --------------------- | | Claude Pro | $20 | 5 | $100 | | Claude Max 5x | $100 | 25 | $20 | | Claude Max 20x | $200 | all 30, $80 to spare | $0 | The planning consequence: the same plan covers a light job thirty times or a heavy one five times, and anything above the credit is ordinary API spend that belongs in the same budget line. Two habits follow. Measure the per-run meter on a handful of real runs before scheduling anything, because token counts from a test prompt rarely match production documents. And watch the boundary as a project grows: automation that starts as your own tool stays inside the credit's terms only while it serves you. The moment other people's requests flow through your plan credentials, it has become the third-party shape Anthropic prohibits, whatever the credit balance says. ## Where this sits next to the OpenAI side The two vendors answered the same question differently, and both answers are deliberate. Anthropic meters programmatic subscription use through these credits and keeps third-party plan routing prohibited. OpenAI includes programmatic Codex inside ChatGPT plan windows, documents `codex exec` for scripts and CI, and documents device-code sign-in for headless machines; third-party use there runs under OpenAI's normal account rules, intended functionality with OpenAI holding the final call. The full structural comparison, windows against credits against meters, is in [Codex vs Claude Code: the cost structure comparison](/blog/codex-vs-claude-code-cost). Our own position follows the policies as written: we do not offer Claude subscription routing, and Claude models pass through ProxyLLM only with your own API-side key. The OpenAI-side automation is what we run on flat plan windows; if that is the bigger line in your budget, the [calculator](/calculator) shows what the flat structure does to it.