AI News — June 23, 2026

Coding agents

• [2026-06-22] Anthropic — Claude Code v2.1.186 added claude mcp login <name> and claude mcp logout <name> to authenticate (or sign out of) an MCP server straight from the CLI without the interactive menu, and made ! bash commands trigger an automatic Claude response by default (opt out with "respondToBashCommands": false). It also added status filtering (press f) to the /workflows detail view, a Skills section to the /plugin Installed tab, and an iterm2 teammate mode, and fixed streaming failures after machine sleep. The non-interactive MCP login is the substantive part — it moves connector auth from a manual menu step into something a setup script or CI job can run unattended. (official)

  Existing users will feel the ! default change before anything else: a quick !git status you fired off just to eyeball the output now draws a full Claude response unless you set respondToBashCommands to false.

  For Software Developers: A container image or CI job that bundles Claude Code can now finish its MCP setup non-interactively — claude mcp login <name> runs as a provisioning step in the Dockerfile or pipeline, so a headless runner comes up already authenticated to its servers instead of stalling on an interactive menu it has no way to answer.

AI cost tracking & telemetry

• [2026-06-17] Anthropic — Workload Identity Federation (WIF) reached general availability on the Claude Platform, letting a workload authenticate to the Claude API with short-lived OIDC tokens minted from an identity provider you already operate — AWS IAM, Google Cloud, GitHub Actions, Kubernetes, SPIFFE, Microsoft Entra ID, or Okta — instead of a long-lived sk-ant-… API key (token lifetime 60–86,400s, default one hour, via the RFC 7523 jwt-bearer grant). It ships alongside service accounts: named non-human identities that each carry their own workspace rate limits, usage attribution, and audit trail, so you can see which workload acted as which account instead of sharing one key across everything. It matters on two fronts — there is no static secret to store, rotate, or leak, and per-workload service accounts turn opaque shared-key spend into usage you can attribute and audit. WIF covers all Claude API endpoints, including the first-party SDKs and Claude Code. (official, docs)

  For many stacks the Claude API key was the last long-lived static secret left in a system that had otherwise gone keyless for its cloud SDKs; WIF lets you retire it and extend the same OIDC trust you already grant AWS or GCP to the model API.

  For Solution Architects: When you’re laying out how a fleet of services calls the Claude API, each one can now hold its own service account federated to the cloud identity it already runs under — the batch job authenticates as itself with a token that expires in an hour, the monthly invoice breaks down per service instead of arriving as one undifferentiated key’s spend, and a compromised node leaks at most a short-lived token rather than the credential every other service shares.

Watch list

• Fable 5 / Mythos 5 restoration — both remain suspended under the June 12 export-control directive, nearly two weeks on with no restoration date; independent status trackers still read “not back” as of June 21–22, and Anthropic’s official notice still commits only to restoring access “as soon as possible.” Today (June 23) the included-plan window ends and usage-credit metering begins — with the model dark, those included days were spent during a blackout no one could use them in. Watch for a restoration date and whether access returns with nationality-based gating attached. (official, plan terms, prior coverage)

  Nationality-based gating is the part with significance beyond this one outage: if a frontier model comes back sold differently to users based on where they’re from, that sets a template for export-controlled model access other vendors would be pressured to follow, long after Fable 5 itself is forgotten.

• Gemini 3.5 Pro (late June) — still a limited Vertex AI enterprise preview with GA targeted for June; days from month’s end, there’s still no model card, public benchmarks, or pricing for the promised 2M-token context and Deep Think mode. (prior coverage)

• xAI Grok V9-Medium — the ~1.5-trillion-parameter, Cursor-trained coding model is reportedly live in the consumer Grok product on X and SuperGrok but still has no API string or model card on x.ai, so nothing testable has shipped for anyone building on it. (prior coverage) (unconfirmed)

• MCP spec finalization (July 28) — the 2026-07-28 release candidate (stateless core, Extensions framework, Tasks, MCP Apps, authorization hardening, formal deprecation policy) is in its validation window; watch for the rest of the SDK support landing before the cutover. (official, prior coverage)

  Watch the stateless core specifically: dropping the requirement that a server hold session state is what lets an MCP server run as a horizontally scaled or serverless backend rather than a sticky long-lived process, so the cutover reshapes how servers get deployed, not just the wire format.