Anthropic Agent SDK Credits Explained: Pricing AI Agents After OpenClaw
Introduction
Anthropic's Agent SDK credits are not just a small perk added to Claude subscriptions. They are a new boundary between interactive Claude usage and programmatic AI-agent usage.
Starting June 15, 2026, eligible paid Claude plans can claim a dedicated monthly credit for Claude Agent SDK usage, claude -p, Claude Code GitHub Actions, and third-party apps that authenticate through the Agent SDK. Ordinary Claude chat, interactive Claude Code, and Claude Cowork continue to use the existing subscription limits.
Key Takeaways
- Agent SDK credits separate programmatic agent usage from ordinary Claude subscription usage
- Pro receives $20, Max 5x receives $100, and Max 20x receives $200 in monthly Agent SDK credit
- Credits are per-user, monthly, non-transferable, and do not roll over
- Claude Code weekly limits are also being raised by 50% through July 13, 2026
- Third-party tools such as OpenClaw can return through Agent SDK authentication, but they draw from this dedicated credit
- The move also looks like a response to OpenAI's Codex-in-ChatGPT-plan strategy
What Are Anthropic Agent SDK Credits?
Agent SDK credits are monthly credits reserved for programmatic Claude usage. They apply only to automation surfaces, not to normal chat usage.
According to Anthropic's help center, the credit covers:
- Claude Agent SDK usage in your own projects
claude -p, the non-interactive Claude Code command- Claude Code GitHub Actions
- third-party apps built on the Agent SDK
It does not cover:
- interactive Claude Code in a terminal or IDE
- Claude conversations on web, desktop, or mobile
- Claude Cowork
- API-key usage through the Claude Developer Platform
The monthly credit amount depends on plan.
| Plan | Monthly Agent SDK credit |
|---|---|
| Pro | $20 |
| Max 5x | $100 |
| Max 20x | $200 |
| Team Standard seats | $20 |
| Team Premium seats | $100 |
| Enterprise usage-based | $20 |
| Enterprise seat-based Premium seats | $200 |
Credits are per-user and reset each billing cycle. Unused credit does not roll over. Once the credit is exhausted, additional usage moves to standard API rates only if extra usage is enabled; otherwise Agent SDK requests stop until the next refresh.
Why Did Anthropic Change the Pricing Model?
The core problem is that subscription pricing and autonomous agent usage do not consume compute in the same way. A chat product assumes a human is reading and responding. An autonomous agent can continue reading files, calling tools, retrying failed steps, and exploring context without a human at the keyboard.
In April 2026, Anthropic restricted third-party agent harnesses such as OpenClaw from drawing on ordinary Claude subscription limits. Reporting from VentureBeat framed the change as a reaction to usage economics: a user paying $20 to $200 per month could run agent workloads that might cost far more at API prices.
Agent SDK credits are the compromise. Anthropic can reopen a path for third-party agents without putting all programmatic usage back into the same subscription bucket as human chat.
This is best understood as a pricing boundary, not simply a free-credit giveaway. Agent workloads return, but their cost is now tracked against a dedicated monthly credit.
Why OpenClaw Became the Flashpoint
OpenClaw is an open-source autonomous agent harness. It lets users connect external model providers and run long-lived agent workflows. That makes it powerful, but also expensive from a model-provider perspective.
The tension is simple: subscriptions are predictable for users, while autonomous agents create variable and sometimes heavy inference load. If an external harness can run indefinitely against a subsidized subscription, the provider absorbs the compute cost.
The new Agent SDK credit system allows tools such as OpenClaw to work through Agent SDK authentication again, but with a clearer cost cap:
- third-party agent usage draws from the Agent SDK credit
- the ordinary Claude chat limit is preserved for interactive use
- usage stops at the monthly credit unless extra usage is enabled
- extra usage is billed at standard API rates
That is why this change can be read both as a reopening and a restriction.
Is This a Response to OpenAI and Codex?
Partly, yes. This is an inference, but the market context points in that direction.
OpenAI's help center says Codex is included with ChatGPT Plus, Pro, Business, and Enterprise/Edu plans. Users can sign in to Codex clients with their ChatGPT account instead of using a separate API key. OpenClaw's OpenAI provider documentation also describes a Codex subscription path using ChatGPT/Codex sign-in.
That creates a competitive comparison:
| Dimension | Anthropic / Claude | OpenAI / Codex |
|---|---|---|
| Subscription agent access | Agent SDK credit bucket | Codex included with ChatGPT plans |
| Third-party agent path | Agent SDK authentication | Codex / ChatGPT subscription auth in OpenClaw docs |
| Overage model | Standard API rates if extra usage is enabled | Plan limits and optional extra credits |
| Pricing philosophy | Separate interactive and programmatic usage | Package Codex as part of the ChatGPT plan experience |
Anthropic likely needed a middle path. Completely blocking third-party agents would alienate power users and developers. Leaving them inside ordinary subscription limits would make costs unpredictable. Agent SDK credits are a controlled reopening.
Why the 50% Claude Code Weekly Limit Increase Matters
The 50% weekly limit increase suggests that Anthropic is not simply tightening access. It is separating automation-heavy usage from ordinary subscription limits while also making interactive Claude Code more competitive.
PC Watch reported, citing the official ClaudeDevs X account, that Claude Code weekly limits are increasing by 50% through July 13, 2026. The change applies to Pro, Max, Team, and seat-based Enterprise users, and works across the CLI, IDE extensions, desktop, and web. This stacks with Anthropic's May 6 announcement that doubled Claude Code's five-hour rate limits and removed peak-hour reductions for Pro and Max accounts.
That combination matters for developer experience. A doubled five-hour window helps intense coding sessions. A 50% higher weekly cap helps longer multi-day projects. Together, they make Claude Code feel less constrained while Agent SDK and claude -p move into a separate credit bucket.
OpenAI is pushing from the other side. Its help center says Codex is included with ChatGPT Plus, Pro, Business, and Enterprise/Edu plans, with temporary access for Free and Go users and 2x rate limits for other plans. OpenAI has also moved Codex to token-based credit pricing and lets Plus/Pro users buy additional credits after hitting included limits.
So the contest is not only about model quality. It is about access architecture.
| Dimension | Anthropic / Claude Code | OpenAI / Codex |
|---|---|---|
| Short-window coding quota | Five-hour limits doubled | Plan-specific Codex limits |
| Weekly coding quota | 50% higher through July 13, 2026 | Codex included in ChatGPT plans |
| Non-interactive agents | Agent SDK credit bucket | Codex/ChatGPT auth and flexible credits |
| Overage model | API rates or stop for Agent SDK usage | Credit purchase and auto top-up |
In that light, the weekly limit increase looks like the other half of the Agent SDK credit change. Anthropic is pulling automated workloads into a priced lane, but giving interactive Claude Code users more headroom so the product does not feel worse than Codex.
What Developers Should Do Now
Developers should treat model access as an architectural choice, not just a login step.
For small personal automation, Agent SDK credits are attractive. They give a predictable monthly allowance, and if extra usage is disabled, they also limit surprise bills.
For long-running OpenClaw workflows, CI automation, or shared production agents, the credit is not the right accounting unit. Those workflows need explicit API billing, budgets, logs, and stop conditions.
A practical decision tree looks like this:
- Personal scripts and experiments: Agent SDK credit
- Interactive Claude Code: existing Claude subscription limit
- Long-running OpenClaw usage: monitor credit burn and plan for API rates
- Team or production automation: Claude Developer Platform API key
- OpenAI-based agent workflows: compare Codex plan limits, optional credits, and API usage
FAQ
Q. Can Agent SDK credits be used for Claude chat?
A. No. Claude web, desktop, mobile, and interactive Claude Code continue to use ordinary subscription limits.
Q. Does this make OpenClaw usable with Claude subscriptions again?
A. It creates a supported path for third-party Agent SDK apps to draw from Agent SDK credits. It does not return unlimited third-party usage to the ordinary subscription bucket.
Q. What happens when the credit runs out?
A. Usage moves to standard API rates only if extra usage is enabled. If it is not enabled, Agent SDK requests pause until the credit refreshes.
Q. Is OpenAI's Codex approach cheaper?
A. Not always. OpenAI packages Codex into ChatGPT plans, while Anthropic separates programmatic usage into a dedicated credit. Which is cheaper depends on workload length, context size, retry behavior, and overage settings.
Summary
Anthropic's Agent SDK credits mark a shift from "one subscription bucket for everything" toward usage-specific accounting for AI agents.
For users, the change makes third-party agent workflows possible again, but less open-ended. For Anthropic, it controls compute exposure while keeping developers in the Claude ecosystem. For the broader market, it shows that the next phase of AI-agent competition will be about pricing architecture as much as model quality.
References
- Use the Claude Agent SDK with your Claude plan
- Higher usage limits for Claude and a compute deal with SpaceX - Anthropic
- PC Watch: Claude Code weekly limits increase by 50%
- VentureBeat: Anthropic reinstates OpenClaw and third-party agent usage
- Using Codex with your ChatGPT plan - OpenAI Help Center
- Codex rate card - OpenAI Help Center
- Using Credits for Flexible Usage in ChatGPT - OpenAI Help Center
- OpenAI - OpenClaw Docs