Latest Anthropic API release notes for June 2026

The June 2026 changes, in order

Newest first. Each row is a public Claude Developer Platform release note.

Late-May groundwork the June notes assume

Several June entries only make sense if you read the three changes that landed days earlier.

The four entries that change your day, if you wrap the agent loop

Most of the list is housekeeping. These four touch how a long-running Claude Code session behaves, what it costs, and where it can run.

How a wrapper actually routes to a new endpoint

The June notes keep adding places Claude can run: AWS endpoints, the existing Bedrock and Vertex paths, corporate proxies. The whole mechanism on the client side is one environment variable. The Anthropic SDK reads ANTHROPIC_BASE_URL, and whatever spawns the agent loop decides what that points at.

Here is the exact behavior in Fazm, which wraps the real Claude Code agent loop in a Mac app. In Desktop/Sources/Chat/ACPBridge.swift (around line 2425), a user-set custom endpoint is validated as an absolute http(s) URL, written into the child process environment, and the bundled key is replaced with a placeholder so your real key never leaves the machine:

if let customEndpoint = Self.validCustomAPIEndpoint(rawCustomEndpoint) {
  env["ANTHROPIC_BASE_URL"] = customEndpoint
  env["FAZM_CUSTOM_API_ENDPOINT"] = "true"
  // A custom endpoint means the user is routing their own model/provider.
  // Never send Fazm's bundled Anthropic key to that proxy.
  env["ANTHROPIC_API_KEY"] = "sk-fazm-custom-endpoint"
}

A malformed value (a missing scheme, a bare localhost:8766, stray text) is rejected and the app falls back to the default Anthropic endpoint, because an invalid base URL otherwise makes the SDK throw on every turn and silently bricks chat. That guard is the difference between "point it at AWS in one field" and "spend an afternoon debugging an empty response."

Why June 15 hits harder than a normal deprecation

Deprecation and retirement are not the same event. Opus 4.1 was deprecated on June 5 with retirement set for August 5, so it keeps serving for now and you get runway. Sonnet 4 and Opus 4 were retired on June 15: after that date a request pinned to those ids does not warn, it fails. Anthropic directs you to Sonnet 4.6 and Opus 4.8 as the replacements.

If your setup hard-codes a model id anywhere (a shell alias, a CI job, a config file, a wrapper's default), that is the line you change. The advantage of a tool that lets you pick the backend model per chat is that a retirement is a one-click swap, not an incident. The same applies to the new Fable 5 and Mythos 5 models from June 9: if you want them, you select them, and the rest of your session is untouched.

1M default context is what makes "don't compact" reasonable

There is a long-running argument about auto-compacting: summarizing the early part of a session to free up room silently drops decisions you made an hour ago. The counter-move is to keep the full history live for the lifetime of the window and never summarize. The obvious objection used to be cost and ceiling: a full history burns context fast.

The late-May Opus 4.8 change moves the ceiling. A 1M token default context (also shared by Fable 5 and Mythos 5) means a long session can carry far more of its own history before anything has to give. That is the technical precondition that turns "never auto-compact" from a stubborn preference into a defensible default. Pair that with mid-conversation system messages that preserve prompt cache hits, also from the May 28 notes, and you can steer a long session without throwing away the cached prefix you already paid for.

If you want the deeper version of this argument, see the companion piece on controlling Claude Code context compaction.

Two billing notes, one quiet win and one that did not happen

The quiet win is June 2: a request returning stop_reason: "refusal" with no generated output is no longer billed. If your agent loop occasionally trips a hard refusal, you were previously paying input tokens for nothing. That stops automatically, no flag required.

The one that did not happen is worth flagging because several guides state it as fact. Anthropic announced a change that would have moved Agent SDK, claude -p, and third-party app usage onto a separate monthly credit pool starting June 15, 2026, then paused it on the day it was due to take effect. So if you bring your own Claude Pro or Max account to a wrapper, that flat-rate path is, as of this writing, unchanged. If a page tells you the credit-pool split is live, check the official notes before you plan around it.

Questions people actually ask about these notes