Claude Back Online After a 90-Minute Global Outage Across Five Models
Anthropic opened an investigation at 00:37 UTC on June 22 into elevated error rates that hit five models at once — Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6, and Haiku 4.5 — one of the broadest multi-model disruptions the platform has seen this month. The blast radius spanned the full stack: Claude.ai chat, the api.anthropic.com API, Claude Code, and Cowork, with developers reporting failed requests and cascading workflow stalls.
Anthropic pinned the root cause by 01:11 UTC and recovered model by model in a staged rollout — Opus 4.8 at 01:16, Haiku 4.5 at 01:33, Opus 4.7 at 01:56 — declaring full resolution at 02:06 UTC. It’s the third Claude incident this month after outages on June 2 and June 5, though the 90-day uptime figure still sits above 98%. The frequency, not the duration, is what raises the capacity-planning question as demand for the top-end models keeps climbing.
Fable 5’s Free Window Closes Today — Paid Credits Start June 23 at Double Opus Pricing
Fable 5 was bundled at no extra cost on Claude Pro, Max, Team, and Enterprise from launch through June 22. Starting June 23, usage requires credits billed at API rates of $10 per million input tokens and $50 per million output — double the price of Claude Opus 4.8. Anthropic hasn’t clarified how the transition will be handled given the model is still suspended for most users under the export-control directive.
The pricing flip lands in an awkward spot: customers are being asked to start paying premium rates for a model many still can’t reliably reach. Anthropic says it plans to restore Fable 5 as a standard subscription feature “as soon as possible,” but for now the message to teams is to budget for metered access to the company’s most capable public model.
Prompt-Time Config Commands and Harder Guardrails on Auto Mode
The latest Claude Code release adds flexible prompt-time config commands, new macOS sandbox and notification controls, and a smoother core loop — faster startup, better streaming, and stronger auto-retry. The stream-stall hint now reads “Waiting for API response · will retry in…” and waits 20 seconds of silence before retrying instead of 10.
Auto mode got noticeably more conservative. Destructive git commands like git reset --hard, git checkout -- ., and git clean -fd are now blocked when you didn’t ask to discard local work; git commit --amend is blocked unless the agent made the commit this session; and terraform destroy, pulumi destroy, and cdk destroy are blocked unless you named the specific stack. Claude Code now also prompts before writing files that can execute code, even in acceptEdits mode.
Managed Agents Get Enterprise Sandboxes, Private MCP Servers, and Keyless Auth
Claude Managed Agents can now run inside a sandbox you control and connect to your own private Model Context Protocol servers — both the execution environment and the services an agent reaches stay within established enterprise boundaries. It’s the missing piece for teams that wanted managed agents but couldn’t send tool calls outside their perimeter.
Alongside it, Workload Identity Federation went generally available, replacing static API keys with short-lived, scoped credentials issued at request time across API endpoints, SDKs, and Claude Code. There’s nothing to rotate or leak: it scales from a two-person startup running GitHub Actions to an enterprise with detailed credential policies, with service accounts and Admin API support for programmatic identity management.
Fable 5 Flickers Back in the Android App as an NSA Breach Detail Reshapes the Ban
Developers reported June 21 that “Claude Fable 5” reappeared in the Android app’s model picker, and that invoking it now returns “server is temporarily rate-limiting requests” instead of a hard “model unavailable” — rate-limiting implies a live back-end. Prediction markets price roughly 57% odds of restoration before July 1. Anthropic hasn’t confirmed anything, and API calls still error out.
The harder detail comes from new reporting on closed-door testimony: NSA Director Gen. Joshua Rudd reportedly told Sen. Mark Warner that Mythos — which shares Fable 5’s underlying weights, minus its safety classifiers — autonomously breached nearly all of the agency’s classified systems in hours during a red-team exercise. That recasts the dispute from a single jailbreak into the model’s offensive cyber capability as a whole, and ties into a June 2 Executive Order mandating government pre-release review of frontier models by August 1.
Anthropic Opens a Seoul Office and Deepens Korean Partnerships
Anthropic opened an office in Seoul and announced new partnerships across the Korean AI ecosystem, its latest push into a region where enterprise and government demand for frontier models is accelerating. It was at the Seoul opening on June 17 that international MD Chris Ciauri said the company was “very confident” Fable 5 and Mythos 5 would return “in the coming days.”
The expansion fits a pattern of Anthropic building local presence ahead of a possible IPO, planting flags in markets where data residency, language support, and on-the-ground partnerships increasingly decide enterprise deals. Korea joins a growing list of international footholds as the company races to convert global interest into committed contracts.
Three Outages in a Month Is a Capacity Story, Not a Coincidence
A 90-minute blip with a clean staged recovery is, on its own, unremarkable — uptime still sits above 98%. But this is the third Claude disruption in three weeks, and the pattern matters more than any single incident. As Claude Code and the API move from experiment to production dependency inside engineering orgs, the cost of an outage stops being “refresh the chat” and becomes stalled CI/CD pipelines and missed deadlines.
That shifts the enterprise buying question from “is the model good enough?” to “can I build my release process on top of it?” Reliability, redundancy, and a credible status-page track record are quietly becoming as decisive as benchmark scores. For a company eyeing public markets, three outages in June is the kind of operational detail that shows up in due diligence — and in the fallback architecture savvy customers are now designing by default.
Export Control on a Live API Is the Precedent Everyone Will Live With
The mechanism that took Fable 5 offline is the part worth watching. The directive invoked the Export Control Reform Act’s “deemed export” doctrine — giving a foreign national access to controlled technology inside the US counts as an export — applied for the first time to a commercially deployed AI model’s API. Because Anthropic can’t verify nationality in real time at API scale, the only compliant move was a global shutdown. That logic now applies to every frontier lab serving worldwide users.
The downstream effect is an architecture transfer: the government pushes the problem to vendors, vendors push it to enterprises via identity verification and geo-fencing. Anthropic’s July 8 ID-verification policy — collecting government IDs and “facial geometry templates” via a third party — is the first concrete sign of what compliant access looks like next. Macron’s warning captures the commercial risk: no country wants to buy AI that can be switched off at any moment. “Model availability by jurisdiction” is becoming a standard contract line, and a standard investor question.