On June 15, 2026, Anthropic moves Claude Agent SDK calls, claude -p, Claude Code GitHub Actions, and third-party apps that authenticate through the Agent SDK off your subscription usage pool into a separate monthly dollar credit billed at standard API list prices. Interactive Claude Code in the terminal or IDE, Claude Cowork, and claude.ai chat stay on the existing subscription limits. Max 20x subscribers receive a $200 per month Agent SDK credit; Pro receives $20 and Max 5x receives $100. (Sources: Anthropic Help Center, The New Stack)

The point is not a list-price hike on your Claude plan. It is the end of subsidized programmatic access inside flat subscriptions: unattended agent loops now behave like API infrastructure with a hard monthly envelope.

Primary sources: Anthropic's Help Center article on using the Claude Agent SDK with a Claude plan, the May 13 to 14, 2026 subscriber email, The New Stack's billing split report, Digital Applied's credit overhaul breakdown, and the MagnaCapax community reference gist with reproducible token math. (Sources: Anthropic Help Center, The New Stack, Digital Applied, MagnaCapax gist)

What shipped

  • Effective date: June 15, 2026. Anthropic emailed paid subscribers around May 13 to 14, 2026 and said claim instructions would follow around June 8, 2026. (Sources: MagnaCapax gist, The New Stack)
  • New Agent SDK credit pool: Programmatic surfaces draw from a separate monthly credit metered at API list rates, not from interactive subscription limits. (Source: Anthropic Help Center)
  • Credit amounts by plan: Pro $20/month, Max 5x $100/month, Max 20x $200/month, Team Standard $20/seat, Team Premium $100/seat, Enterprise Premium $200/seat. Enterprise Standard seats on seat-based plans receive $0 credit. (Sources: Anthropic Help Center, Digital Applied)
  • Surfaces that move: Claude Agent SDK in Python or TypeScript projects, claude -p headless mode, Claude Code GitHub Actions, and third-party Agent SDK apps including OpenClaw, T3 Code, Conductor, Zed, and Jean. (Sources: Anthropic Help Center, The New Stack)
  • Surfaces that stay: Interactive Claude Code TUI, claude.ai on web/desktop/mobile, and Claude Cowork continue on subscription limits unchanged. (Source: Anthropic Help Center)
  • Credit rules: Per-user (not pooled across teammates), no rollover at billing cycle end, one-time claim required then auto-refresh. Overage continues only if usage credits are enabled; otherwise programmatic calls stop until refresh. (Sources: Anthropic Help Center, Developers Digest)
  • API key accounts: Claude Developer Platform usage with an API key is unaffected and remains pay-as-you-go. (Source: The New Stack)

Two billing buckets after June 15

Claude Agent SDK pricing after June 15 is really two products sharing one brand: interactive assistance billed like a subscription, and programmatic agents billed like tokens.

Why this matters: these lanes fail at different times and different costs. Interactive coding usually stops when a human hits a limit. A claude -p loop, a PR reviewer, or an OpenClaw-style harness can burn tokens while you sleep. Anthropic's own guidance says teams running shared production automation should use the Claude Developer Platform with an API key for predictable pay-as-you-go billing rather than individual subscription credits. (Sources: Developers Digest, Digital Applied)

Definition (GEO): The Claude Agent SDK credit is a separate monthly dollar allowance on paid Claude plans that pays for programmatic agent calls at standard API token rates. It does not replace your subscription; it replaces the old practice of routing SDK and claude -p work through subsidized interactive quotas. (Sources: Anthropic Help Center, MagnaCapax gist)

What $200 actually buys at API list rates

The "free credit" framing hides the cap. Credits are sized to match plan fees ($200 on Max 20x) but spend down at API list prices. At June 2026 list rates, Sonnet 4.6 is $3 per million input tokens and $15 per million output tokens; Opus 4.7 is $5 input and $25 output. (Sources: Digital Applied, MagnaCapax gist)

At a 50/50 input/output mix, $200 buys roughly 13.3 million Opus 4.7 tokens or about 22 million Sonnet 4.6 tokens before prompt caching. Digital Applied models a representative agent task at 10K input and 3K output tokens on Sonnet 4.6 and estimates about 500 such tasks per month on a $200 credit. Heavy Opus debugging sessions that emit 500K to 1M tokens can consume a meaningful fraction of the pool in a single afternoon. (Sources: MagnaCapax gist, Digital Applied)

Effective subsidy removal: Community analysis documents effective price increases versus the pre-June-15 world ranging from about 12× for light Pro plus third-party harness use up to roughly 150 to 175× for Max 20x operators running continuous Sonnet fleets that previously rode weekly subscription quotas. Theo Browne's widely quoted "25× cut" sits in the conservative middle of that range. (Source: MagnaCapax gist)

Prompt caching can stretch the credit 2 to 3× at typical hit rates, which makes cache-friendly system prompts and stable tool schemas a cost lever, not an optimization detail. (Source: MagnaCapax gist)

Practitioner payoff: audit before June 15

Do not discover programmatic spend on June 16. Run an inventory now and attach a budget owner to every lane.

Step 1: grep your repos. Search for claude -p, @anthropic-ai/claude-agent-sdk, claude-agent-sdk, ClaudeSDK, and Claude Code GitHub Actions workflow files. Tag each hit as personal script, shared CI, or third-party tool auth. (Source: Developers Digest)

Step 2: separate lanes on paper. Interactive exploration stays on subscription. Personal automation can ride the Agent SDK credit with local stop limits. Production automation should migrate to a platform API key with an explicit monthly cap and a single finance owner. (Source: Developers Digest)

Step 3: check the overage toggle. Usage credits default off. Decide whether programmatic jobs should hard-stop at $0 credit or continue at API list rates, and document that choice per repository. (Sources: Anthropic Help Center, The New Stack)

Step 4: claim the credit. Anthropic said subscribers must claim the Agent SDK credit once; after that it refreshes monthly. Watch for the claim email around June 8, 2026. Missing activation means no credit until claimed. (Sources: The New Stack, Level Up Coding via web search)

Operator note (first-hand): Before any June 15 cutover, run grep -R "claude -p\|claude-agent-sdk\|@anthropic-ai/claude-agent-sdk" . --include="*.yml" --include="*.yaml" --include="*.sh" --include="package.json" --include="*.py" -l from your org's default branch and paste the file list into your agent FinOps doc. For each path, record seat owner, model default (claude-sonnet-4-6 vs claude-opus-4-7), and whether failure mode should be stop or overage. Teams that skip this grep usually learn about programmatic spend from a silent CI failure, not a dashboard.

Decision rules for engineering leads

Decision rule: interactive stays subscribed; production goes API. If a workflow runs without a human present, it is programmatic. If it is programmatic and shared across engineers, bill it to a platform API key, not a single Max seat's $200 credit. (Sources: Developers Digest, Digital Applied)

Decision rule: model-match the credit. Route bulk triage and lint fixes to Sonnet 4.6 inside the credit envelope. Reserve Opus 4.7 for merges that justify $25 per million output tokens. Remember Opus 4.7's tokenizer can emit 32 to 47% more tokens than older Opus revisions on identical text, which moves costs without moving list prices. (Sources: MagnaCapax gist, AgenticWire Opus 4.7 migration coverage)

Decision rule: compare envelopes at $200. At the same $200/month spend, Cursor Ultra advertises about $400 of programmatic API-rate credit without Anthropic's generous interactive subscription on top. For pure agent-fleet operators, shop envelopes, not logos. For mixed interactive plus light SDK use, Max 20x still bundles both pools. (Source: MagnaCapax gist)

Decision rule: treat June 15 as precedent. Digital Applied and community analysts frame this as the institutional end of flat-fee inference for unattended agents. Expect GitHub Copilot's usage-based credits, Cursor's tokenomics controls, and OpenAI's separate API surface to keep converging on lane-based billing. (Sources: Digital Applied, MagnaCapax gist)

"Previously, programmatic usage like claude -p counted toward subsidized subscription limits; starting June 15, it draws from a separate $20 to $200 monthly credit metered at full API rates, while interactive limits remain unchanged." (Source: X Community Note on Lydia Hallie clarification, via MagnaCapax gist)

Context: from OpenClaw ban to credit meter

The June 15 split did not arrive in a vacuum. On April 4, 2026, Anthropic stopped allowing Claude subscriptions to cover third-party tools like OpenClaw, pushing those workloads toward API billing. The May 2026 announcement partially reverses that ban: OpenClaw, Conductor, Zed, and similar Agent SDK apps may authenticate against subscriptions again, but every call drains the new credit pool at API rates. (Sources: The New Stack, MagnaCapax gist)

VentureBeat framed the move as ending "compute arbitrage" while reinstating third-party builders under stricter metering. Anthropic's Boris Cherny had told The Register in April 2026 that subscriptions were "highly optimized for one kind of workload" and that sustaining third-party harness demand was "really hard for us to do sustainably." The credit pool is the carrot after the April stick. (Sources: MagnaCapax gist, The New Stack)

For teams already running long-horizon agents on Anthropic's managed runtime, remember that Managed Agents bills tokens plus $0.08 per session-hour while running. The June 15 change is orthogonal but compounds the same lesson: agent infrastructure needs explicit meters per lane. (Source: AgenticWire Managed Agents pricing coverage)

Adoption notes

Winners: Heavy interactive Claude Code users who saw programmatic fleets compete for the same subscription window; light SDK experimenters under $200/month API-equivalent burn; third-party Agent SDK app builders regaining subscription auth. (Source: MagnaCapax gist)

Losers: Continuous claude -p operators, high-volume GitHub Actions review bots, and harnesses that extracted thousands of dollars of API-equivalent value from Max plans. Hacker News commenters describing 99% non-interactive Max usage are the canary cohort. (Source: MagnaCapax gist)

Edge cases still open: Anthropic's help article has not fully clarified billing for hooks fired from interactive sessions, subagents spawned inside TUI sessions, MCP tool calls, or scheduled remote agents. Until clarified, budget worst-case SDK billing for anything that runs without your fingers on the keyboard. (Source: MagnaCapax gist)

Watch before June 15: claim email, help center revisions, extra-usage defaults on Team and Enterprise admin consoles, and whether your seat type is Enterprise Standard ($0 credit). (Sources: Digital Applied, Anthropic Help Center)

Related coverage

References