TL;DR (June 2026): Two Anthropic changes land on the same day, June 15, 2026. First, Claude Agent SDK usage, headless claude -p, Claude Code GitHub Actions, and third-party apps authenticating through the Agent SDK leave your subscription's usage pool and start consuming a separate monthly dollar credit ($20 on Pro, $100 on Max 5x, $200 on Max 20x) billed at standard API rates, per user, no rollover, and you must claim it once before June 15. Second, claude-opus-4-20250514 and claude-sonnet-4-20250514 are retired the same day: API requests to them fail. Interactive Claude Code in the terminal and claude.ai chat are unchanged. Here is the checklist.
If the Tokenpocalypse was the month AI coding tools stopped pretending to be flat-rate subscriptions, June 15 is the day Anthropic finishes the thought. The pattern is exactly the one GitHub ran on June 1: the usage that was quietly subsidized inside a flat plan gets carved out, denominated in dollars, and metered at real rates. This time the carve-out is agents. And because Anthropic picked the same date to retire two models, a lot of teams are going to hit two unrelated failure modes in the same deploy window and misdiagnose one as the other.
Both changes are verifiable against primary sources: the billing split was announced through Anthropic's help center in mid-May, and the retirements have been on Anthropic's model deprecations page since April 14. Here is what each one actually does, the math, and the five things to do this week.
Change 1: agents leave the subscription
Until June 15, a Pro or Max subscription was a single pool: interactive Claude Code sessions, chat, and programmatic Agent SDK workloads all drew from the same usage limits. That made a Max plan a remarkably cheap way to run automation, and plenty of people noticed: CI pipelines, scheduled agents, and third-party tools all riding a $100/month plan at what was effectively a deep discount to API rates.
From June 15, the following usage no longer draws from subscription limits:
- Claude Agent SDK workloads (programmatic agents, personal projects, products built on the SDK)
claude -p, the headless / non-interactive mode- Claude Code GitHub Actions
- Third-party apps that authenticate through the Agent SDK with your subscription
Instead, that usage consumes a separate monthly credit, billed at standard API list rates:
| Plan | Monthly agent credit | Roughly what it buys (Sonnet 4.6) | Roughly what it buys (Haiku 4.5) |
|---|---|---|---|
| Pro ($20/mo) | $20 | ~2-3M tokens | ~8-10M tokens |
| Max 5x ($100/mo) | $100 | ~13-17M tokens | ~40-50M tokens |
| Max 20x ($200/mo) | $200 | ~27-33M tokens | ~80-100M tokens |
Token estimates assume Sonnet 4.6 at $3/$15 and Haiku 4.5 at $1/$5 per million input/output tokens, with a 3:1 input-to-output mix; a 1:1 mix lands at the lower end. Opus 4.7 ($5/$25) burns roughly 3x faster than Sonnet. Treat these as planning numbers, not quotas; your actual mix decides.
Four mechanics matter more than the headline numbers:
- You must claim the credit, once, before June 15. It is an opt-in through your Claude account, then it refreshes automatically each cycle. Miss the claim and your SDK workloads have no pool to draw from on cutover day.
- No rollover. Unused credit expires with the billing cycle, the same use-it-or-lose-it shape as GitHub's AI Credits.
- Per user, not pooled. A team of five Max seats is five separate $100 credits, not a shared $500.
- Exhaustion behavior is a setting. When the credit runs out, SDK requests either fail until the next cycle, or overflow to usage credits at standard API rates, but only if you have explicitly enabled usage credits. Whichever you want, choose it deliberately before June 15, because the default you have not thought about is the one that will surprise you: silent failures in CI, or an uncapped overflow bill.
What does not change: interactive Claude Code in the terminal and claude.ai web/desktop chat stay on subscription limits. This split is aimed at exactly one thing, the gap between human-interactive use (which subscriptions were priced for) and automated agent use (which was being subsidized). It is the same lesson Uber and Microsoft learned the expensive way, covered in the budget-bomb story: agent workloads scale like infrastructure, not like people. For the full cost picture now that the SDK is on its own meter, including the June 22 Fable 5 change, see what Claude Code actually costs in 2026.
Change 2: Claude 4 Opus and Sonnet 4 are retired the same day
Separately, and easy to confuse with the billing change because of the shared date, two models hit their retirement date on June 15, per Anthropic's deprecations page (deprecated April 14, 2026):
| Retiring June 15, 2026 | Recommended replacement | Pricing note |
|---|---|---|
claude-opus-4-20250514 | claude-opus-4-8 | Claude Opus 4 listed at $15/$75; the current Opus tier (4.7) lists at $5/$25, roughly two-thirds cheaper |
claude-sonnet-4-20250514 | claude-sonnet-4-6 | Same $3/$15 list price, newer model |
Requests to retired models fail. Not degrade, not redirect: fail. Any pipeline still pinning claude-opus-4-20250514 or claude-sonnet-4-20250514 starts erroring on June 15. And the next one is already scheduled: claude-opus-4-1-20250805 was deprecated on June 5 and retires August 5, 2026.
There is also a quieter breakage worth knowing about while you migrate: on Opus 4.7 and later, the temperature, top_p, and top_k parameters are deprecated and return a 400 error when set to non-default values. A naive find-and-replace of the model string can therefore still break a request that has worked for years. Migration is a small project, not a string swap.
The good news: migrating off Claude Opus 4 is one of the rare forced migrations that cuts your bill. At list prices, the same Opus-class workload costs roughly a third of what it did, and our Claude API pricing and limits guide has the full current rate card.
The same-day trap
Because both changes land on June 15, a team that did nothing will see two different failures at once: SDK workloads erroring because there is no claimed credit pool, and pinned Claude 4 calls erroring because the model is gone. The symptoms look similar from a failing pipeline (requests stop working), and the fixes are completely unrelated. If your June 15 incident channel lights up, triage them separately: check the model ID in the error first, then check the credit state.
The checklist (this week, in order)
- Claim the agent credit in your Claude account. One click, one time, before June 15. Do it for every seat that runs SDK workloads.
- Audit your model IDs. Grep your codebase and config for
claude-opus-4-2025,claude-sonnet-4-2025, and while you are there,claude-opus-4-1(August 5). Anthropic's Console has a usage export (Usage page, Export button) that breaks usage down by API key and model, which is the fastest way to find forgotten consumers. - Audit your headless usage at API rates. Everything that runs through the Agent SDK or
claude -pnow meters in dollars. If you do not know what your CI and scheduled agents consume in tokens per month, you cannot know whether $20 or $200 covers it. This is exactly the per-workload metering problem covered in the FinOps playbook. - Set the exhaustion behavior deliberately. Enable usage-credit overflow if a stalled pipeline is worse than a bigger bill; leave it off if a surprise bill is worse than a paused agent. Either is defensible, defaulting into one by inattention is not.
- Tell whoever owns CI. Claude Code GitHub Actions are in scope. The person who finds out on June 15 that the nightly agent run died is rarely the person who read the announcement.
The honest take
This is the Tokenpocalypse pattern repeating with unusual clarity. The subscription was never priced for automated agents; it was priced for a human at a keyboard, and the gap between those two costs was a subsidy that providers across the category are now withdrawing one by one. GitHub did it June 1 with AI Credits. Anthropic is doing it June 15 with the SDK carve-out. The direction is not in question anymore.
What is worth saying plainly: the metered version is more honest, and for disciplined teams it is often cheaper, because the flat plan's "unlimited" was already gone, replaced by opaque rate limits. A dollar-denominated credit you can measure beats a vibes-denominated limit you cannot. But it only beats it if you actually measure: per-workload token attribution, burn-rate against the monthly credit, and an alert before exhaustion rather than a failure after it. That is a metering-layer problem, not a tool-choice problem, and it is the same conclusion every one of these transitions keeps landing on. The teams that meter will treat June 15 as a non-event. The teams that do not will discover their automation footprint the way Uber did: from the bill.