On June 15, 2026, the Claude Agent SDK credits system goes live. Anthropic is splitting programmatic Claude usage away from your Pro, Max, Team, or Enterprise subscription limits and putting it on a separate monthly credit pool, billed at Anthropic API rates once you opt in. The change covers the Claude Agent SDK in your own projects, the claude -p headless command in Claude Code, Claude Code GitHub Actions, and third-party apps that authenticate against the Agent SDK. Interactive use of Claude.ai and Claude Code in the terminal is unaffected. This is the most consequential pricing change Anthropic has shipped in 2026 for anyone running Claude inside agents, CI pipelines, or embedded products, and the action items for builders are time-sensitive.
Quick Answer
- From June 15, 2026, the Claude Agent SDK and
claude -pno longer draw from your Claude plan's usage limits. Anthropic's official help-center article confirms the date and scope. - Each paid plan gets a one-time opt-in monthly Agent SDK credit: Pro $20, Max 5x $100, Max 20x $200, Team Standard $20, Team Premium $100, Enterprise usage-based $20 or seat-based Premium $200.
- Credits are billed at standard Anthropic API per-token rates. When the monthly credit runs out, Agent SDK requests either bill into "usage credits" (if enabled) or simply stop until the next cycle.
- Claude Code used interactively in a terminal or IDE, plus Claude conversations on web, desktop, or mobile, keep drawing from your normal subscription limits exactly as before.
- Teams running Claude Code GitHub Actions, scheduled agents, or any non-interactive automation should claim credits early, decide whether to enable overage, and plan budgets in API tokens, not subscription seats.
What changes on June 15, 2026
Anthropic's official help-center entry, titled Use the Claude Agent SDK with your Claude plan, gives the cleanest description we have. Starting June 15, 2026, Claude Agent SDK usage and claude -p calls no longer count toward your Claude plan's regular usage limits. Instead, each paid plan becomes eligible for a dedicated monthly Agent SDK credit. You claim it once through your Claude account, after which it refreshes automatically each billing cycle. Until you claim, programmatic calls still draw against your normal subscription, but on June 15 the line moves.
The credit is denominated in dollars, not tokens or "messages", and is consumed at the same per-model rates Anthropic publishes for the API. There is no separate Agent SDK tariff. Claude Fable 5, Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5 all bill at their standard API list prices when called through the Agent SDK or claude -p. The new model context, including the Claude Fable 5 launch on June 9, 2026, makes the per-token math even more important than it was a month ago. See our Claude Fable 5 launch breakdown for the model side of the equation.
What counts as programmatic usage (and what does not)
The split is precise. Programmatic usage, which now consumes the Agent SDK credit, is:
- Claude Agent SDK used in your own projects (Python or TypeScript).
claude -p, the non-interactive headless mode of Claude Code, often used in scripts and CI.- Claude Code GitHub Actions (PR reviews, scheduled jobs, comment triggers).
- Third-party apps that authenticate against the Agent SDK with your Claude credentials, including community frameworks like OpenClaw.
Interactive usage, which still draws from your standard subscription, is:
- Claude.ai conversations on web, desktop, and mobile apps.
- Claude Code in the terminal or IDE when you are typing prompts and the model is replying to you, with no
-pflag. - Claude Cowork, the shared workspace.
The single most common confusion is Claude Code itself: the same binary acts as interactive Claude Code (subscription pool) when you type to it, and as headless Claude Code (Agent SDK credit pool) when you call claude -p. The flag is the dividing line. If you script Claude Code from a shell loop, a cron job, or a Makefile target, you are on the Agent SDK side of the wall starting June 15. For background on how the SDK fits into Anthropic's stack, see our Claude Agent SDK 2026 guide.
Claude Agent SDK credits, plan by plan
The per-plan amounts come directly from Anthropic's help-center article and align with what InfoWorld and The New Stack reported. They are designed to mirror the cash price of the subscription, with one important wrinkle for Enterprise.
| Plan | Monthly Agent SDK credit | What it buys at API rates (illustrative) |
|---|---|---|
| Pro ($20/mo) | $20 | ≈ 1.3M input + output tokens on Opus 4.8, ≈ 5M on Sonnet 4.6, ≈ 25M on Haiku 4.5 |
| Max 5x ($100/mo) | $100 | ≈ 6.5M on Opus, ≈ 25M on Sonnet, ≈ 130M on Haiku |
| Max 20x ($200/mo) | $200 | ≈ 13M on Opus, ≈ 50M on Sonnet, ≈ 260M on Haiku |
| Team Standard | $20 per seat | Same as Pro per seat |
| Team Premium | $100 per seat | Same as Max 5x per seat |
| Enterprise (usage-based) | $20 per seat | Same as Pro per seat |
| Enterprise (seat-based Premium) | $200 per seat | Same as Max 20x per seat |
Token estimates are rough, derived from Anthropic's published API list prices and an assumed 50/50 input-to-output ratio. Your actual cost depends on the model, prompt size, output size, and any caching you have configured. The point of the table is order of magnitude: a $20 monthly credit on Opus 4.8 covers roughly one or two solid headless coding sessions, while the same $20 on Haiku 4.5 covers tens of millions of tokens for triage and routing work.
What happens when credits run out
Anthropic spells this out clearly: when your monthly Agent SDK credit is depleted, additional Agent SDK and claude -p requests fall through to your account's usage credits at standard API rates. If you have not enabled usage credits, those requests are rejected until the credit refreshes at the start of your next billing cycle.
In practice this gives builders two distinct postures:
- Strict ceiling: Leave usage credits off. The Agent SDK budget is exactly your plan's monthly credit. When you run out, everything programmatic stops, predictably and visibly. Best for solo developers, side projects, and anyone who wants no billing surprises.
- Auto-overflow: Turn on usage credits with a top-up. Programmatic work keeps flowing at API list prices after the monthly credit is gone. Best for agencies, SaaS embedders, and production pipelines where a hard cutoff in the middle of a CI run is worse than a higher bill.
There is no automatic conversion from interactive subscription quota into Agent SDK credits or vice versa. The two pools are independent.
Who actually pays more, and who barely notices
Three builder profiles capture most of the impact.
The interactive developer. You open Claude Code, type prompts, and let it run. You do not use claude -p. You do not run Claude Code in CI. Your bill is unchanged. You can ignore June 15.
The agent operator. You run Claude Code in headless mode, GitHub Actions on every PR, or your own Agent SDK loops on a server. Before June 15 your subscription absorbed all of that work under a generous fair-use ceiling. After June 15 every headless call eats either your $20 to $200 Agent SDK credit or your usage-credit balance at API rates. If you were running heavy headless workloads on a Pro plan, your effective monthly Claude budget for that work just collapsed from "very high" to "$20 of API tokens", and you should plan accordingly. This is the cohort the change is aimed at.
The SaaS embedder. You ship a product that calls the Agent SDK on behalf of your customers. You were already on a metered model: either direct Anthropic API keys or a Pro subscription stretched across many users. After June 15 the right mental model is "Anthropic API at list price". Two consequences: you need a way to pass that variable cost into your own pricing without exposing it to end users, and you need a predictable production surface for the apps your customers actually ship, one that does not flicker when credit pools refresh. For embedders thinking about this, the Calendly link at the bottom of this article is the right next step.
For broader context on Anthropic's current SaaS and team economics, see our Claude Code pricing 2026 breakdown and the related Anthropic Managed Agents 2026 playbook.
A practical playbook for the next 7 days
If you ship code or product that touches the Claude Agent SDK, work through this checklist between now and June 15.
- Claim your credit on day one. Log into your Claude account and complete the one-time opt-in. It refreshes automatically after that.
- Decide the overage stance. Pick strict ceiling or auto-overflow before June 15. Do not let the default decide for you under pressure.
- Audit every place that calls Claude headlessly. Grep your repo for
claude -p, any Agent SDK imports, and GitHub Actions workflows that invoke Claude. Note the model used in each. - Migrate routine tasks to cheaper models. Triage, routing, classification, and small refactors that ran on Opus by default should move to Haiku 4.5 or Sonnet 4.6. The Agent SDK credit will stretch 5x to 20x further.
- Add prompt caching and turn down output verbosity. Both move tokens down the bill. The Claude Code MCP tutorial covers the integration patterns where this matters most.
- Set a billing alert. Whether you opt for the hard cutoff or auto-overflow, set a cost alert at 70% and 100% of your expected monthly burn.
- Plan the production side. If your customers will eventually see this cost passed through, decide now whether you will charge per credit, per app, or per seat. Embedders building on top of Totalum get a stable production envelope that does not change on June 15. See our guide on how to embed an AI app builder in your SaaS via API.
Where Totalum fits in this change
Totalum is the AI app builder for humans and for agents. It is not a Claude Code subscription, and it is not affected by the June 15 Agent SDK credit split in the same way a claude -p user is. When a Claude Code session, an Agent SDK loop, or a custom agent talks to Totalum, it is calling Totalum's own API and MCP to build, deploy, and run a real Next.js application with built-in auth, payments, database, file storage, and a custom domain.
That separation matters for two audiences in particular:
- Software agencies running client work through agent fleets get a deployment target whose price does not flip on a quiet Anthropic billing toggle. The Agent SDK side of the workflow still moves to per-token billing, but the production side stays steady.
- SaaS companies embedding an AI builder get a clean way to expose vibe-coding to end users without exposing them to Anthropic's split-pool billing or your own Anthropic credit ceiling. You meter your customers however you want, and Totalum runs the build, deploy, and hosting layer underneath.
If either of these patterns sounds like your team, the Calendly link below is the fastest way to talk through the integration.
For a broader picture of where Claude Code is heading as a platform, see our Code with Claude Tokyo 2026 recap, the Claude Code hooks playbook, and the Claude Code subagents guide.
FAQ
When exactly does the Claude Agent SDK credits change take effect?
The new policy starts on June 15, 2026. Until then, Claude Agent SDK and claude -p calls keep drawing from your existing Claude plan's usage limits. On June 15 they move to the separate monthly Agent SDK credit, and you complete a one-time opt-in to claim it. Subsequent months refresh automatically.
Does this change Claude API pricing or Claude.ai pricing?
No. Anthropic's API per-token rates and the public subscription prices ($20 Pro, $100 Max 5x, $200 Max 20x) did not change. The change is structural: the pool that programmatic usage draws from is now separate from the pool that interactive chat and interactive Claude Code draw from.
Are interactive Claude Code sessions in the terminal affected?
No. Interactive Claude Code, where you type prompts and the model replies in your terminal or IDE, still draws from your normal subscription limits exactly as before. Only claude -p (headless) and Agent SDK calls move to the new credit pool. The flag is the dividing line.
What happens if my Agent SDK credit runs out mid-month?
If you have usage credits enabled on your Claude account, additional Agent SDK requests continue and bill at standard Anthropic API per-token rates. If usage credits are disabled, those requests are rejected until the next monthly cycle refreshes the Agent SDK credit. There is no auto-borrow from your interactive subscription quota.
Do Claude Code GitHub Actions count as Agent SDK usage?
Yes. Claude Code GitHub Actions runs are non-interactive by definition, so they consume Agent SDK credits starting June 15, 2026, exactly like claude -p and the Agent SDK itself. Teams running review bots, scheduled jobs, or comment-triggered automation should expect those workflows to draw from the new credit pool.
How do I keep my Agent SDK credit lasting longer?
Move routine work (triage, routing, classification, small refactors) from Opus to Haiku 4.5 or Sonnet 4.6, turn on prompt caching, cap output token budgets per call, and avoid one-shot Opus calls for tasks a smaller model handles well. A Pro plan's $20 credit can buy roughly one full Opus session or tens of millions of tokens of Haiku work, so the model choice often matters more than the prompt.
Should I switch to direct Anthropic API keys instead?
If your usage is high enough that the Agent SDK credit will not cover even a few days, direct API keys with billing alerts are often cleaner. If you are at lower volume or want a single Claude account billing surface for both interactive and programmatic work, claim the Agent SDK credit, set the overage toggle deliberately, and stay on the subscription. SaaS companies embedding Claude for their end users should treat this as standard API economics and price accordingly.
Ready to build with Totalum?
The Claude Agent SDK credits change is a strong signal that programmatic Claude usage is moving to a fully metered model. If you run an agency, ship client apps from agent fleets, or want to embed an AI app builder into your SaaS without exposing your customers to Anthropic's split-pool billing, Totalum's API and MCP let you keep a predictable production envelope underneath whatever model billing structure Anthropic ships next.
For software agencies and SaaS embedders, book a 30-minute call to see Totalum live: calendly.com/totalum/30min.
For solo builders and AI agent operators who want to plug Claude Code into a real production stack, start free at www.totalum.app.
Sources
- Anthropic Help Center: Use the Claude Agent SDK with your Claude plan.
- Anthropic API pricing: platform.claude.com/docs/en/about-claude/pricing.
- The New Stack: Anthropic splits billing again: Agent SDK gets separate credit pools.
- InfoWorld: Anthropic puts Claude agents on a meter across its subscriptions.
- Anthropic Claude Platform release notes: platform.claude.com/docs/en/release-notes/api.