Dynamic Workflows and Auto Mode Land on Bedrock, Vertex, and Foundry
The agent-orchestration story is no longer single-cloud. Auto mode is now available on Amazon Bedrock, Google Vertex AI, and Microsoft Foundry for Opus 4.7 and Opus 4.8, and dynamic workflows — ask Claude to build a workflow and it orchestrates work across tens to hundreds of background agents — are in research preview across the Claude Code CLI, Desktop, and the VS Code extension for Max, Team, and Enterprise plans, plus the Claude API and all three hyperscaler platforms.
The practical read for enterprise buyers: the highest-leverage agentic features no longer require leaving your existing cloud commitment. For teams in regulated industries that procure AI through Bedrock or Vertex for data-residency reasons, parity across platforms removes the last “but it’s only on the first-party API” objection — and it lands four days before a Tokyo keynote with a dedicated Claude Platform track.
June 15 Cutover Is 8 Days Out — Retired Model IDs Start Failing Monday Week
Eight days remain until June 15, when Claude Sonnet 4 (claude-sonnet-4-20250514) and Claude Opus 4 (claude-opus-4-20250514) retire on the Claude API. Anything still calling those strings begins failing that day. The same date brings the Agent SDK credit split — Pro gets $20/month in agent credit, Max 5x gets $100, Max 20x gets $200, all expiring monthly — so any unattended pipeline needs a spend cap before the meter starts.
The checklist is unchanged and worth running this weekend: grep your codebase for pinned model IDs, repoint to current models (Opus 4.8 is the frontier as of late May), claim your credit from the account-email instructions, and cap anything running headless. Yesterday’s fallbackModel setting doubles as cheap insurance — name a current model as fallback so a stray retired ID degrades gracefully instead of hard-failing.
Claude Code v2.1.167: A Quiet Sunday Reliability Drop
One day after the headline fallbackModel release, v2.1.167 landed with bug fixes and reliability improvements rather than new surface area — the kind of incremental hardening that has defined the whole week. There’s no flashy feature here, which is itself the point: after a sub-agent outage and a fast cadence of resilience changes, Anthropic is tightening screws, not adding knobs.
If you upgraded for fallbackModel yesterday, take the point release too — back-to-back drops this close together usually mean a fix for something the prior build surfaced. For fleet-managed installs, this is exactly the scenario this week’s version-pinning controls were built for: pin a known-good range, then promote 2.1.167 once you’ve smoke-tested it.
The Connector Directory Tops 418 — and the Same Integrations Now Run as Agents
Claude’s verified-connector directory has grown past 418 integrations across 30 categories, spanning revenue tools, creative software, developer platforms, and now everyday consumer apps. The more interesting structural change is that Managed Agents, launched in April, pair directly with these MCP connectors — meaning the same integrations powering interactive chats can run unattended as long-running background workflows with no extra setup.
For builders, the takeaway is that the connector is becoming the unit of capability. Wire up an MCP server once and it works in chat, in Claude Code, and inside an autonomous agent. That collapses the old gap between “a thing Claude can talk about” and “a thing Claude can do on a schedule,” and it’s the quiet reason the directory’s headcount matters more than any single integration.
Code w/ Claude Tokyo Is 3 Days Out — In-Person Closed, Livestream Open
The June 10 Tokyo program is set: check-in and breakfast, a product keynote, then breakouts and workshops across three tracks — Research, Claude Platform, and Claude Code — with demos and office hours running 10:30 AM to 8:00 PM and live English/Japanese interpretation throughout. In-person applications have closed; virtual registration via the livestream page is still open, and the June 11 Extended day targets independent developers and early-stage founders.
The keynote material practically assembled itself this week: version pinning, fallbackModel, glob deny rules, the cross-session security fix, and now dynamic workflows on every major cloud. The open questions for Tuesday are whether Anthropic uses the stage to finally deliver the missing outage RCA, names a Japan systems-integrator partner to match the new Services Track, and says anything firmer about the Mythos public-release timeline.
Strava and Abacor Push MCP Into Consumer and Back-Office Territory
Two June 1 launches show how wide the connector net is now cast. Strava’s MCP connector gives subscribers OAuth read-only access to per-second heart-rate and pace streams, GPS routes, cycling power, and club metadata — conversational analysis of years of training data, and the clearest mass-consumer MCP integration of the year. On the opposite end, Abacor’s connector exposes meeting data, summaries, and action items, and is designed to chain with Gmail, QuickBooks, Xero, and Microsoft 365.
The pattern worth noting: connectors are arriving from both the consumer-fitness and the SMB-accounting ends of the market at once, not just from the usual enterprise-software vendors. That breadth is what turns a directory count into a network effect — the more domains that ship a connector, the more reasons any given user has to keep Claude as the hub rather than a point tool.
The S-1 Math: A $47B Run-Rate, a $965B Round, and a Model That Writes Its Own Code
With the confidential S-1 in SEC review, the financial picture underneath it is now public: a $65B Series H at a $965B post-money valuation — ahead of OpenAI’s reported $852B — an annualized revenue run-rate that recently crossed $47B, and reporting that points to a $1.75–1.8T IPO target and a raise as large as $75B. Whatever the final pricing, those are numbers that make this the public-market test case for the entire AI-infrastructure thesis.
The detail that ties the week together: Anthropic says Claude now writes more than 80% of its own code, up from under 10% in early 2025, at roughly human parity. That is the flywheel the S-1 is implicitly selling — a model that accelerates its own development feeds the run-rate that justifies the valuation. The skeptic’s footnote is the one this brief keeps returning to: a company pitching itself as infrastructure still owes the market a written post-mortem for the June 2 outage. Reliability narratives and IPO narratives are the same narrative now.
Mythos Is Still “Coming Weeks” Out — and That Phrase Is Doing a Lot of Work
Anthropic continues to say Mythos-class models reach all customers “in the coming weeks,” with the alignment safeguards already shipped inside Opus 4.8 — whose misaligned-behavior rates Anthropic puts near Mythos Preview’s. The remaining gate isn’t capability; it’s cyber-misuse prevention for a model that can autonomously run complex network attacks and surface critical software flaws.
Watch how the release gets paced. The most likely shape is a Cyber Verification Program that meters access by who can be trusted with offensively-capable tooling rather than a flip-the-switch general launch — the same logic behind scaling Mythos to critical-infrastructure defenders first. For anyone in aerospace, defense, or industrial systems, the timeline that matters isn’t the model’s; it’s the verification program’s.