Three Chatbots, One Bad Afternoon
Claude went down yesterday at roughly 9:41am ET and stayed degraded for three hours and six minutes. Elevated error rates, stalled responses and failed requests hit the web app and the API together, across Mythos, Fable, Sonnet 5 and Opus 5. Downdetector logged more than 6,000 reports at peak, and the single largest complaint category was Claude Code — which tracks, because a coding agent mid-session does not degrade gracefully the way a chat window does.
The genuinely strange part is the company it kept. ChatGPT and Grok went down in the same window. Three independent frontier providers, three separate infrastructure stacks, one afternoon. No shared root cause has been published, and coincidence is a perfectly ordinary explanation when everyone is scaling into the same handful of clouds and CDNs at the same time. But if your architecture’s fallback plan for Claude being unavailable was “route to another provider,” yesterday was the day that plan got tested and did not hold.
Everything is resolved and reading operational. The practical takeaway is not about Anthropic’s uptime, which remains good. It is that multi-provider failover is not the same thing as independent failure domains, and most teams have been quietly assuming it is.
Enterprise Frontier Safeguards Puts the Logs in Your Bucket
Enterprise Frontier Safeguards (EFS) is Anthropic’s attempt to end a trade-off enterprises have been complaining about for two years: you could have zero data retention, or you could have misuse detection, but not both, because detection needs history and ZDR deletes it. EFS resolves that by moving the history somewhere else — specifically into the customer’s own cloud account, whether that is Amazon S3, Azure Blob Storage or Google Cloud Storage, under keys the customer controls.
It was designed with more than 100 customers spanning financial services, healthcare, manufacturing, telecom, law, retail and the public sector, and it will surface across Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google’s Agent Platform and Microsoft Foundry. Rollout is phased, starting later this fall.
One caveat is getting hammered in the coverage and deserves the airtime: Anthropic retains read access to the stored data, for the misuse detection the feature exists to enable. That is internally consistent — you cannot detect cross-session abuse in a bucket you cannot read — but it means “your data, your cloud, your keys” describes custody, not isolation. If your compliance story depends on the distinction, read the terms rather than the headline.
$65B Raised, and Roughly Fifteen Gigawatts Contracted
The capital picture underneath everything else in this brief: Anthropic’s Series H raised $65 billion at a $965 billion post-money valuation, led by Altimeter, Dragoneer, Greenoaks and Sequoia, against a run-rate revenue that crossed $47 billion. Growth since the Series G in February has come from enterprise, which is why the last two weeks read like an enterprise roadmap rather than a research one.
Almost all of that money has a destination, and it is compute. Amazon for up to five gigawatts of new capacity. Google and Broadcom for five gigawatts of next-generation TPU capacity. SpaceX for GPU access in Colossus 1 and Colossus 2. Microsoft and NVIDIA for Claude scaling on Azure. Those are not vendor relationships; at that scale they are closer to joint infrastructure programs with multi-year build schedules attached.
Two things follow. First, Anthropic has deliberately refused to standardize on a single silicon vendor or a single cloud — TPUs, NVIDIA GPUs and Trainium all in the mix. Second, the $47B against $965B gap means the valuation is priced on where the compute buildout gets you in three years, not on this year’s revenue. Both facts make the customer-controlled infrastructure work elsewhere in today’s brief look less like generosity and more like clearing the last procurement objection standing between Anthropic and the enterprise budget it has already borrowed against.
A Fullscreen Diff Panel, and MCP Servers Your Org Controls
Claude Code’s current build adds the feature people have been asking for in every thread about reviewing agent output: a fullscreen diff panel that opens beside the conversation and shows uncommitted changes as Claude edits them, toggled with /diff. Watching a diff accumulate in real time is a materially different review posture than scrolling back through tool calls afterward.
The administratively significant addition is managedMcpServers, a managed setting that lets an organization provision HTTP/SSE MCP servers to every user centrally. That converts MCP configuration from a per-developer .json file into something security teams can actually govern. Alongside it: --permission-prompts none for unattended headless hosts, --json on claude plugin validate for machine-readable CI output, and recognition of GitLab merge requests, which now render as MR !N and accept /code-review --comment findings via glab mr note.
Rounding it out: prompt-cache miss diagnostics now appear in /cost and the status line’s prompt_cache field, with a likely cause attached. Given what caching costs at agent scale, being told why you missed is worth more than being told that you missed.
Self-Hosted Claude Code Sessions, on Runners You Provision
Self-hosted environments for Claude Code are in public beta for Team and Enterprise organizations, moving cloud sessions off Anthropic’s infrastructure and onto servers inside the customer’s network. Sessions still start from web, mobile, desktop or a routine — but they execute next to your internal services, toolchains and security controls. Repository checkouts, build artifacts, secrets and any files the session touches stay on machines you provisioned.
Two runner modes. Fixed keeps a set number of runners with sessions distributed across them. On-demand watches the queue, spins a runner up when work arrives and shuts it down when the work finishes, so capacity tracks demand rather than headcount. It is off by default, and notably not available to organizations already on ZDR — a constraint that EFS looks designed to eventually dissolve.
Anthropic is unusually direct about the cost: this is for teams whose network, tooling or compliance requirements demand it, and who plan to staff engineers to own setup and ongoing maintenance. That is the right framing. Self-hosting an agent runner is not a checkbox, it is a service you now operate.
Persistent REPL State Lands Across Seven SDKs
The Python, TypeScript, Go, Java, Ruby, PHP and C# SDKs now all support code_execution_20260120 — the code execution tool version that adds REPL state persistence. Variables, imports and loaded data survive between calls instead of every execution starting from a blank interpreter. It is also the minimum version required for programmatic tool calling, so this is a floor, not an option, if you are heading that direction.
Separately, MCP tunnels (research preview) moved their management API out of the Admin API and into the Claude API, behind the anthropic-beta: mcp-tunnels-2026-06-22 header and the workspace:manage_tunnels WIF scope. Small migration, but it is a breaking path change — worth grepping for if you built tooling against the old endpoint.
The through-line on both: statefulness is being pushed back into the execution layer after the 2026-07-28 MCP spec deliberately pulled it out of the protocol. That is the correct separation. The transport should be stateless and cheap to deploy; the sandbox is where memory belongs.
40,000 Firms Applied. 10,000 Consultants Are Certified.
The Claude Partner Network, launched in March with $100 million in 2026 funding, has pulled in more than 40,000 firm applications and produced over 10,000 consultants holding a Claude certification. The anchor tenants are the obvious ones — Accenture, Deloitte, Cognizant and Infosys — and a Services Track and Partner Hub have since been added on top.
Those numbers describe a channel, not a community. Ten thousand certified consultants is roughly ten thousand people whose billable rate depends on Claude being the answer, embedded inside the firms that write enterprise architecture recommendations. The Accenture multi-year partnership is explicitly framed around moving enterprises from AI pilots to production, which is the exact stage where a vendor gets locked in or swapped out.
Pair this with Claude Academy opening free courseware earlier this week and the shape is clear. Anthropic is building the same two-tier distribution machine that AWS and Salesforce built before it: free credentials at the bottom to standardize the labor pool, paid systems integrators at the top to close the deals. It is unglamorous and it works.
80% of Anthropic’s Own Production Code Is Written by Claude
A figure worth sitting with: more than 80% of the code merged into Anthropic’s production codebase in May was authored by Claude, not by humans. The company reports an 8x increase in code shipped per engineer per quarter against its 2021–2025 baseline.
Treat the number carefully. “Authored by” is not “unreviewed,” and lines merged is a famously poor proxy for value delivered — an agent that writes verbose, well-tested, boilerplate-heavy code will dominate that metric without necessarily dominating the outcome. The 8x shipping figure is the more meaningful one, and even that is measured inside a company with unusual model access, unusual tooling and unusual incentives to make the number look good.
What it is genuinely useful for is calibration. If the org with the best possible conditions lands at 80% authored and 8x throughput, that is roughly the ceiling of the current generation — not the baseline your team should feel behind for missing. The gap between that number and yours is mostly test coverage, review capacity and codebase legibility, which are the things worth investing in before the next model lands.
The Trust Boundary Is Moving, and Yesterday Explained Why
Read today’s three biggest items in sequence and they tell one story. Enterprise Frontier Safeguards moves your monitoring data into your cloud account. Self-hosted environments move your agent execution onto your runners. And yesterday, for three hours and six minutes, the shared layer everyone depends on stopped working — at the same time as two competitors’. Those are not three unrelated stories. They are the same argument arriving from two directions.
The direction of travel for the last two years has been maximal centralization: send everything to the provider, let the provider handle inference, storage, monitoring, execution, all of it. That was rational while capability was scarce and the deployment surface was a chat box. It stops being rational once the thing running is an agent with repository access, secrets and a build pipeline. The blast radius changed, and infrastructure follows blast radius.
What is notable is that Anthropic is building the decentralization itself rather than resisting it. EFS and self-hosted runners both hand real control back to the customer, and both cost Anthropic something — less telemetry, less operational leverage, more support surface. The trade is that the enterprise deals with the largest compliance requirements become closeable, and the customer’s exit cost goes up rather than down, because migrating an integrated self-hosted runner fleet is much harder than switching an API endpoint. Handing over the keys and deepening the lock-in turn out to be the same move.
The honest caveat is the one The Register raised on EFS: the data lives in your bucket, but Anthropic keeps read access, because misuse detection requires it. Every architecture like this has that seam somewhere, and the useful question is never “is this really zero retention” — it is which specific party can read what, under which conditions, and who audits it. Anthropic deserves credit for shipping the mechanism and for being reasonably legible about the limits. Buyers deserve to read the contract anyway.
The financing frames the rest. $65 billion raised, $965 billion post-money, $47 billion run-rate, and something like fifteen gigawatts contracted across Amazon, Google, Broadcom, SpaceX and Microsoft. A company that has borrowed that heavily against three-year enterprise revenue does not ship customer-controlled log storage and self-hosted runners out of principle. It ships them because the largest remaining objection in the room is procurement, not capability — and every gigawatt on order has to be paid for by someone whose security team currently says no.