Anthropic paused the Claude Agent SDK credits program on June 15, 2026, hours before it was supposed to take effect. The official help-center article now opens with a notice: "We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed: Claude Agent SDK, claude -p, and third-party app usage still draw from your subscription's usage limits." This page covers what was supposed to ship, what is actually live today, the simultaneous Claude 4 model retirements that did go through, and the portable-architecture move that protects builders from the next reversal.
Quick Answer
- Anthropic paused the Claude Agent SDK credits program on June 15, 2026, the day it was supposed to go live.
- The carve-out that would have moved Claude Agent SDK,
claude -p, Claude Code GitHub Actions, and third-party SDK apps out of subscription rate-limits into a separate $20 to $200 monthly credit pool is on hold. - For now, SDK and headless
claude -pusage still counts against your Claude plan's normal usage limits. No new effective date has been announced. - The other June 15 event did ship: Anthropic retired
claude-sonnet-4-20250514andclaude-opus-4-20250514. Pinned API calls to those IDs now return errors. Migrate to Sonnet 4.6 and Opus 4.8. - Two pricing reversals in 30 days (Fable 5 suspension on June 12, credits paused on June 15) is the signal. The strategic answer is portable architecture: build with abstractions that survive vendor turbulence.
What was supposed to happen on June 15
Anthropic announced the Agent SDK credit split on May 14, 2026 with a clear effective date of June 15. We covered the original plan in detail in our Claude Agent SDK credits in 2026 breakdown. The short version:
- Claude Agent SDK,
claude -p, Claude Code GitHub Actions, and third-party Agent SDK applications would no longer count toward your Claude plan's interactive rate limits. - They would instead draw from a new, separate monthly credit pool, billed at standard API list prices.
- Allocations: Pro $20, Max 5x $100, Max 20x $200, Team Standard $20, Team Premium $100, Enterprise (usage-based) $20, Enterprise (seat-based Premium) $200.
- Mechanic was a one-time opt-in claim. Credits would not roll over. Overflow would either fail or fall through to usage credits at standard API rates, depending on your account setting.
A wide swath of the developer press, including Codersera, The New Stack, and UsageBox, spent the past month explaining the math. Build CI teams quietly migrated to the opt-in flow. Indie hackers ran spreadsheets to break-even.
Then on June 15, the official Anthropic help article was rewritten.
Claude Agent SDK credits paused: what is actually live today
The current support.claude.com article on Agent SDK plan usage starts with the pause notice quoted in the intro. The body of the page still describes the carve-out mechanics, but the previously announced monthly credit "isn't available." There is no new effective date. There is no replacement program announced. Anthropic says they will share an update before anything takes effect.
Read literally, today's state is the pre-May-14 status quo: SDK and headless work draws against the same usage limits as interactive Claude.ai or Claude Code sessions. If you were carefully throttling your CI to avoid blowing through the upcoming $20 Pro credit, you can revert. If you were waiting to opt in, there is nothing to opt in to.
The relevant operational consequences:
- CI pipelines that depend on
claude -por Claude Code GitHub Actions continue to share quota with human developers on the same plan. The cliff some teams were planning around has not arrived. - Third-party SDK apps, which were originally banned a few months back, then restored as part of the credit-pool plan, are still authorized against subscription limits per the current help-article text.
- The pricing risk has shifted from "definitely on June 15" to "could ship at any time with short notice." That second state is in some ways harder to plan around than a fixed date.
The other June 15 event that did ship
While the credit split was paused, Anthropic's Claude 4 model retirements went through on schedule. The platform release notes for June 15, 2026 are explicit:
> We've retired the Claude Sonnet 4 model (claude-sonnet-4-20250514) and the Claude Opus 4 model (claude-opus-4-20250514). All requests to these models will now return an error.
The recommended replacements are claude-sonnet-4-6 and claude-opus-4-8. There is no grace period. Pinned-version production traffic against the deprecated IDs is failing as of yesterday.
This is the part of the June 15 narrative that is easy to miss because the credits pause sucked up the press cycle. If you have not run a code search across your repos and infrastructure for the literal strings claude-sonnet-4-20250514 and claude-opus-4-20250514, do it today. Our Claude 4 deprecation migration guide walks through the one-line swap for production deployments.
Planned vs paused: the comparison table
| Dimension | Planned for June 15 | Actually live June 15 to 16, 2026 |
|---|---|---|
| Agent SDK billing source | Separate monthly credit pool | Same Claude plan usage limits as interactive |
| Pro tier programmatic budget | $20/month credit | Shared usage limits |
| Max 5x tier programmatic budget | $100/month credit | Shared usage limits |
| Max 20x tier programmatic budget | $200/month credit | Shared usage limits |
claude -p billing |
Separate credit | Subscription limits |
| Claude Code GitHub Actions billing | Separate credit | Subscription limits |
| Third-party SDK apps | Authorized via credit pool | Authorized via subscription |
| Overflow behavior | Configurable: hard-fail or usage-credit overflow | N/A, not on the metered split |
| Opt-in step | One-time claim required before June 15 | No opt-in available; program paused |
| Effective date | June 15, 2026 | Postponed indefinitely. "Anthropic will share an update before anything takes effect." |
Claude Sonnet 4 (pinned 20250514) |
Retired June 15 | Retired. API returns errors. |
Claude Opus 4 (pinned 20250514) |
Retired June 15 | Retired. API returns errors. |
The pause is partial, not full. The pricing change paused, the model retirements did not.
Why this matters for AI app builders
The substance of the credit-split plan was fine. The strategic problem is the volatility of the announcement: a multi-week, well-documented, broadly communicated pricing change reversed itself the day it was supposed to ship. Combined with the Claude Fable 5 and Mythos 5 suspension on June 12, that is two materially disruptive Anthropic policy moves inside one week.
A few practical takeaways:
- Don't hardcode pricing assumptions into automated systems. If your CI shutdown logic, cost-allocation report, or break-even calculator depends on the $20/$100/$200 monthly credit numbers shipping on a particular date, those assumptions are now wrong.
- Don't pin to specific Claude model strings if you can avoid it. The Sonnet 4 and Opus 4 retirements show that Anthropic will burn deprecated model IDs on schedule. Use latest-stable aliases (
claude-sonnet-4-6,claude-opus-4-8) where your application can tolerate auto-upgrades. - Have a non-Anthropic fallback path. Kimi K2.7-Code, DeepSeek V4 Pro, and the Claude family with Sonnet 4.6 + Opus 4.8 form three independent vendor lanes. Routing logic that can swap one for another in a config change buys you survivability through suspensions, deprecations, and pricing reversals.
- Watch the help-center article, not the press cycle. The pause notice is the canonical source of truth right now. Subscribe to the article via whatever change-tracking workflow you have. The next move from Anthropic will land there first.
What to do this week
A short, opinionated checklist for teams shipping production work on Claude:
- Grep your repos for
claude-sonnet-4-20250514andclaude-opus-4-20250514. Replace withclaude-sonnet-4-6andclaude-opus-4-8. Redeploy. This is non-optional, the old IDs are dead. - Pull back any "claim your credit before June 15" reminders from internal Slack channels and onboarding docs. There is nothing to claim.
- Audit your CI cost. Now that the metered split is on hold, the old throttles you put in place for it are spending budget you do not need to spend. Re-tune to interactive-shared limits and watch for the inevitable next change.
- Document your Claude vendor-risk posture in one paragraph that you actually re-read quarterly. Include: which model strings you depend on, your fallback model if Anthropic flips a switch, and the maximum dollar exposure you have if pricing changes overnight.
- If you build with embedded AI app builders for customers (agencies, SaaS embedders), test your provider abstraction. Can you swap the underlying coding agent without changing your customer-facing surface? That is the property that protects you the most.
The portable-architecture move with Totalum
Totalum is the most powerful AI app builder for humans and for agents. It generates real Next.js + TotalumSDK applications with built-in auth, payments, database, file storage, AI integrations, deployment, and custom domains. The same builder is reachable via API and MCP, so software agencies and SaaS embedders can use it as their build engine.
Two properties matter for the conversation above:
- You own the output. A Totalum-generated application is a deployable codebase with its own database and your domain. If Anthropic changes a pricing rule next week, your application keeps running on whatever Claude model string you pinned, alongside any non-Claude lane you wired in. There is no captive runtime layer that breaks if a single vendor moves.
- Provider abstraction is in the box. The AI integrations layer in a Totalum app can talk to Claude, OpenAI, xAI, Moonshot, DeepSeek, and other providers, with model selection in your application config. The cost of switching when a vendor reverses a pricing decision is a config change, not a rewrite.
That is the practical answer to a week with two reversals: build the architecture so the next reversal is a fifteen-minute config change.
Ready to build something portable?
If you are an individual developer or an early-stage team, start free at totalum.app and ship a real application this week. The first project includes auth, payments, database, file storage, AI calls, and a custom domain out of the box.
If you are a software agency building client work on Claude, or a SaaS company embedding an AI app builder into your product, book a 30-minute discovery call and we will walk through how Totalum's API and MCP layer fits into your delivery model.
FAQ
Q: Did Anthropic cancel the Claude Agent SDK credits program?
A: No, they paused it. The official help-center article says "we're pausing the changes" and "Anthropic is working to update the plan to better support how users build with Claude subscriptions." There is no new effective date and no replacement program announced. Treat the program as on hold rather than dead.
Q: Do I still need to claim my Pro $20, Max 5x $100, or Max 20x $200 credit?
A: No. The opt-in claim was tied to the now-paused June 15 effective date. The credit "isn't available," per Anthropic's current help article. Remove any internal reminders that ask team members to claim the credit before June 15.
Q: How are Claude Agent SDK and claude -p calls billed today?
A: They draw from your Claude plan's normal subscription usage limits, just as they did before May 14, 2026. Heavy CI workloads will eat into the same quota that interactive Claude Code sessions use. Plan capacity accordingly.
Q: Did the Claude Sonnet 4 and Opus 4 retirements also get paused?
A: No. Those went through as planned on June 15, 2026. Requests to the pinned IDs claude-sonnet-4-20250514 and claude-opus-4-20250514 now return errors. Migrate to claude-sonnet-4-6 and claude-opus-4-8. See our Claude 4 deprecation migration guide for the swap.
Q: When will the Agent SDK credits program ship?
A: Unknown. Anthropic says they "will share an update before anything takes effect." There is no public timeline. Watch the Claude Agent SDK plan-usage help article for the next version.
Q: How should I protect my application against the next pricing reversal?
A: Use latest-stable model aliases, not pinned dates, where you can. Wire a non-Anthropic fallback provider into your routing layer (Kimi K2.7-Code or DeepSeek V4 Pro are reasonable Claude-class options). Build with an abstraction layer like Totalum so swapping providers is a config change. Treat Anthropic policy moves as recurring events, not one-offs.
Q: Is this related to the Fable 5 and Mythos 5 suspension?
A: Not directly. The Fable 5 / Mythos 5 suspension on June 12, 2026 was driven by a US government export directive. The Agent SDK credits pause is a commercial-policy reversal with no stated regulatory tie. The common thread is that both interrupt builder plans on short notice, which reinforces the case for portable architecture.
Sources
- Anthropic Help Center: Use the Claude Agent SDK with your Claude plan (canonical source of the pause notice)
- Anthropic Claude Platform API release notes (June 15, 2026 retirements)
- Codersera: Anthropic's June 15 Billing Change
- The New Stack: Anthropic splits billing again
- UsageBox: Anthropic's June 15 Double Hit
- Anthropic Model Deprecations