Anthropic's latest API release notes, June 2026

What shipped, in order

Six items from the June 2026 platform notes, each with the one line that decides whether it touches your day.

1. Claude Fable 5 generally available

   Jun 9

   Anthropic's most capable widely released model, with a 1M input context and up to 128K output tokens. Always-on thinking, no assistant prefill, raw reasoning never returned, and a 30-day data-retention requirement. Priced above the Opus tier at $10 / $50 per million tokens.

2. Server-side fallbacks parameter (beta)

   Jun

   With header server-side-fallback-2026-06-01, a request that a safety classifier declines is re-run on a named fallback model inside the same call, with credit repricing applied automatically. Claude API and Claude Platform on AWS only; rejected on the Batches API; not on Bedrock, Vertex, or Foundry.

3. Refusals with no output are no longer billed

   Jun

   A request that returns stop_reason: "refusal" before any tokens are generated costs nothing. A mid-stream refusal still bills the partial output. Either way, clients must read the stop reason rather than blindly index content[0].

4. MCP tunnels (research preview)

   Jun

   Connect to MCP servers that live inside a private network, without exposing them publicly. Relevant if your tools sit behind a corporate boundary.

5. Self-hosted Managed Agent sandboxes

   Jun

   Tool execution can now run in your own infrastructure instead of Anthropic's, while the agent loop stays on the orchestration layer. Managed Agents also gained live tool/MCP config updates on an active session, and 100K+ token tool outputs now spill to a file in the sandbox.

6. Rate Limits API

   Jun

   Administrators can now query the rate limits configured for an organization and its workspaces programmatically, rather than inferring them from 429 headers.

Fable 5, by the numbers

The headline model of the month. The context window and output ceiling matter for long agentic runs; the price is why it is not the default Opus-tier upgrade.

What this looks like from a desktop Claude Code client

Most write-ups of an API changelog stop at the wire format. The question that is missing is: if you are not calling /v1/messages yourself, when do any of these reach you? Fazm is a useful case to reason about because it does not reimplement Claude Code. It wraps the real agent loop over the Agent Client Protocol. The bridge bundles @agentclientprotocol/claude-agent-acp, pinned to 0.29.2 in acp-bridge/package.json, and you connect your own Claude Pro or Max account.

The consequence: model availability and API behavior come from Anthropic's side and your plan, not from a model list compiled into the app. When Fable 5 became available on your account, it was reachable through the same loop with no Fazm release. The same is true of the quieter changes, the refusal stop reason and the fallbacks parameter, because they live in the request and response shapes that the agent loop already speaks.

The seam where a stop reason becomes a UI decision

When Anthropic adds or changes a stop reason or an error shape, the client has to decide what the user sees. This is the part that cannot be lifted from a generic changelog, so here is the actual code. Fazm's bridge ships a single helper, classifyApiFailure, in acp-bridge/src/api-failure.ts. It sorts a thrown session error into three buckets: overloaded, credit, or other.

The load-bearing line is if (httpStatus === 529) sitting before the credit check. Anthropic returns 529 for overloaded_error, but the Claude Agent SDK tags those api_retry events with error: "rate_limit". A client that reads only the error type treats an Anthropic outage as the user running out of credit, restarts the subprocess, and aborts every other open chat with a billing message. That is exactly what happened in production on 2026-05-14, and it is why the status code is checked first. The fuller postmortem is in the May 2026 529 write-up.

June's changes land in this same place. The new unbilled refusal stop reason is a fourth kind of upstream condition: not an overload, not a credit problem, and not a generic error to dump verbatim. It is a deliberate decline with no charge. The right behavior for a desktop client is to show it as a terminal-but-clean turn, the same way this classifier keeps a transient 529 from masquerading as a billing event. The server-side fallbacks parameter is the API-side counterpart: where your own code can opt a refused request into a retry on a second model, a wrapper that runs the real loop inherits that behavior through the request shape rather than reimplementing it.

If you do call the API directly

For completeness, the shape of the new fallbacks parameter, since it is the one June change most likely to need a code edit on your side. Branch on stop_reason before reading content, and opt into a fallback so a single classifier decline does not fail the request outright.

The header string must be exactly server-side-fallback-2026-06-01; the parameter is rejected on the Message Batches API and unavailable on Bedrock, Vertex, and Foundry, where the client-side refusal-fallback middleware is the recommended substitute. Stop reason values for a Fable 5 refusal include cyber, bio, and reasoning_extraction; see the Anthropic error reference.