Update (2026-06-22): Anthropic paused this on the day it was due. The split was scheduled for 2026-06-15 but Anthropic paused it that day — nothing changed for now; claude -p, the Agent SDK, and GitHub Actions still draw from your normal subscription limits (the-decoder, digitalapplied). It remains officially announced and could still return — so treat this page as preparation in case it comes back, not a live emergency.
Under the announced change, Anthropic would move programmatic Claude usage off your subscription pool onto a separate, dollar-metered credit pool. Interactive use is unchanged. It was due 2026-06-15 but was paused that day (see above) — here's who would be affected and what to do if it returns.
claude -p, Claude Code GitHub Actions, or a third-party agent, that usage leaves your normal subscription limit on 2026-06-15 and starts drawing from a fixed monthly credit billed at full API rates, with no rollover. If you only use Claude Code interactively in your terminal (and claude.ai / Cowork), nothing changes.
| Pool | What it covers | How it's billed |
|---|---|---|
| Interactive (unchanged) | claude.ai web/desktop/mobile, Claude Code used interactively in the terminal, Claude Cowork | Your normal subscription limits, exactly as before |
| Agent SDK credit (new) | Agent SDK, claude -p, Claude Code GitHub Actions, third-party agent apps | A fixed monthly dollar credit at full API list prices, no rollover |
| Plan | New Agent SDK credit / month |
|---|---|
| Pro | $20 |
| Max 5× | $100 |
| Max 20× | $200 |
| Team — Standard seat | $20 |
| Team — Premium seat | $100 |
| Enterprise — usage-based | $20 |
| Enterprise — Premium seat | $200 |
Metered at full API list prices with no rollover. Because it's billed at API rates rather than the subscription's effective rate, heavy programmatic users can see the same workload cost noticeably more once it draws from this pool — which is why estimating your programmatic usage now matters. Note: on seat-based Enterprise plans, Standard seats can't claim the Agent SDK credit; only Premium seats can.
claude -p "…" in scripts or cronThe honest question is how much of your usage actually crosses to the new pool. Two reliable ways, in order of certainty:
claude -p / claude --print, the Agent SDK, or GitHub Actions is programmatic by definition, so this is the most certain check. Grep your scripts, cron, and CI:
grep -rn "claude -p\|claude --print" ~ 2>/dev/null and check your GitHub workflows for anthropics/claude-code-action. Every hit is Pool-2 usage after June 15.CLAUDE_CODE_ENABLE_TELEMETRY=1 and OTEL_METRICS_INCLUDE_ENTRYPOINT=true, then group your token/cost metrics by the app.entrypoint attribute — the programmatic entrypoints are your Pool-2 volume. This is the method Anthropic documents (monitoring docs); teams can also use the Claude Code Analytics API.Tip: once you know your monthly programmatic token volume, price it at the full API list rates (Opus 4.5–4.8 input $5 / output $25 / cached read $0.50 per million; Sonnet 4.6 $3 / $15; Haiku 4.5 $1 / $5) to see whether it fits inside your plan's credit. Source: Anthropic pricing.
claude -p / CI / Agent SDK workload already costs more than your plan's credit at API rates, plan to reduce it or budget the overflow.The billing split isn't the only thing landing on 2026-06-15. Two API-level breaking changes share the same date, and they hit code, not your plan. If you have model IDs pinned in scripts, CI, or config, your calls can start failing that day.
Anthropic deprecated these on 2026-04-14 with a 2026-06-15 retirement date. After June 15, requests to a retired model ID return an error.
| Retiring ID (stops working) | Replacement |
|---|---|
claude-opus-4-20250514 | claude-opus-4-8 |
claude-sonnet-4-20250514 | claude-sonnet-4-6 |
Find pinned IDs before they break:
grep -rn "claude-opus-4-20250514\|claude-sonnet-4-20250514" . \
--include=*.{js,ts,py,json,yml,yaml,toml,sh,env}Also export your API usage by model from the Claude Console Usage page (Export → CSV) to catch IDs your code sets dynamically.
temperature / top_p / top_k now 400 on Opus 4.7+If your "fix" for the retirement is to bump the model to Opus 4.7 or 4.8, note that setting any of temperature, top_p, or top_k to a non-default value returns a 400 error on Opus 4.7 and later. Remove those parameters and steer behavior through the prompt instead.
Sources: Anthropic model deprecations page (verify against your own claim email and the live docs, since dates can change).
If you're affected, the lever is spending fewer tokens per programmatic run so the new credit lasts. The biggest wins are usually prompt-cache hygiene, model selection per task, trimming context that's re-sent every turn, and catching cost-spike patterns (large tool outputs, hook stderr re-injection, duplicated context). A field reference of 13 Claude Code cost-spike patterns (free) covers the common ones with a defense for each.
No. Interactive terminal use keeps drawing from your normal subscription limits. Only programmatic usage (Agent SDK, claude -p, GitHub Actions, third-party agents) moves to the new credit pool.
They draw from a fixed monthly credit ($20 Pro / $100 Max 5× / $200 Max 20×) metered at full API list prices, no rollover. Estimate your current programmatic token spend at API rates to see whether you'll exceed it. If you do, you either reduce usage or enable usage-credit overflow (billed at API rates).
By default, programmatic requests past the credit are rejected. If you enable "usage credits," they instead bill at API rates so your automation keeps running. Choose based on whether budget or uptime matters more for that workload.
No action required for interactive use. It's worth claiming the credit anyway in case you start using agents later.
Not for interactive use. For programmatic use it can effectively raise cost, because that usage now bills at full API list prices from a capped credit rather than drawing from the subscription's effective rate. Heavy automation users feel it most.
Anthropic's official classification names only four programmatic paths (Agent SDK, claude -p, GitHub Actions, third-party SDK apps). It does not explicitly classify claude remote-control, MCP servers, or in-session sub-agents, so treat any confident claim about them with caution. The structural reading is that they follow the entrypoint that launched them: an MCP call or sub-agent inside an interactive terminal session stays interactive (Pool 1), while the same inside a claude -p / SDK run is programmatic (Pool 2). remote-control is reported to invoke claude --print internally, which would put it on the programmatic side — but Anthropic hasn't confirmed this. Don't guess: measure with the app.entrypoint attribute above on your own usage.
Independent summary by an operator running Claude Code 800+ hours, maintainer of cc-safe-setup (free MIT hooks). Announced by Anthropic on 2026-05-13, effective 2026-06-15 — always confirm the exact terms against Anthropic's own claim email and docs, since plans can change. For a deep, field-tested guide to cutting Claude Code token cost (the lever for surviving the new credit pool), see the Token Book (Claude Code Token Optimization Guide, English). Not affiliated with Anthropic.