Anthropic Signs a $45 Billion Compute Deal With Nscale
Anthropic has agreed to spend roughly $45 billion over six years renting AI cloud capacity from Nscale, the British infrastructure company, drawing primarily on Nscale’s flagship West Virginia data center development. The commitment works out to approximately 460 megawatts of power, built on Nvidia’s Vera Rubin systems — the multi-chip architecture that sits at the current frontier of accelerator design.
The timing detail is the one worth holding onto: the capacity does not start serving Claude until late 2027. Vera Rubin silicon begins coming online late next year, and Anthropic’s slice of it lands after that. This is not capacity that relieves any pressure on today’s rate limits or today’s incident rate. It is a bet placed on demand roughly eighteen months out.
This is the latest and largest in a run of compute agreements Anthropic has signed over the past eight months as it scales against OpenAI, and it arrives with the company reportedly weighing an October listing. Neither company has published a formal announcement; the figures come from Bloomberg and CNBC sourcing rather than a filing, so treat the precise number as reported rather than confirmed.
Claude Mythos 5 Goes Into Claude Security — Plus a $35M Defender Fund
Anthropic has put Claude Mythos 5, its most capable and most access-restricted model, inside Claude Security, the vulnerability scanning product currently in public beta for Claude Enterprise customers. Point it at a repository and it returns findings classified by CWE, with severity, a confidence rating and suggested patches. There is no separate SKU and no new contract — usage bills as ordinary token consumption.
The framing here is deliberate and unusual. Enterprises get what the model finds; they do not get the model. Mythos-class direct access stays locked behind the Cyber Verification Program, which today gives vetted organizations reduced safeguards on Opus and Sonnet and is due to widen in the coming weeks to broader dual-use capabilities, with Mythos-tier access following later.
Alongside it, Anthropic announced the Defender Advantage Fund (0xDAF), $35 million in Claude credits directed at open-source software security. Priority goes to remediating live vulnerabilities in widely used projects, scalable scan-and-patch automation, and work that defends against whole classes of attack rather than single CVEs. Anthropic is also working with security vendors to embed the model into tools defenders already run.
The Watermark Is Live on Everything Shipped Since August 2
Worth restating because it applies to output you are generating right now: every Anthropic model released after August 2 automatically watermarks generated text and files. The technique biases Claude’s word choices in a statistically detectable pattern — based on Google DeepMind’s SynthID-Text — and it survives copy-paste.
The driver is the EU AI Act Transparency Code, which took effect August 2 and requires machine-readable marking of AI-generated or AI-edited content. Anthropic is one of roughly 190 signatories to the associated code of practice.
The friction is concentrated in one place, and it is not the place most coverage expected. Lawyers, academics and writers who use Claude to edit work that is substantially their own now find that work carrying an AI marker. If your team runs Claude as a copy editor over human drafts, that is a policy question to settle internally before someone else runs a detector on it.
Claude Code Now Writes Its Own Feedback Reports
Announced this morning: Claude Code files its own feedback reports. Three triggers — when a tool keeps failing, when Claude notices it made a mistake, and when you call it out. The signal that used to die in your terminal now goes upstream on its own.
This is a smaller feature than the Agent tool that shipped in 2.1.246, and probably a more consequential one for release quality. The historical failure mode with agentic coding tools is not that they break loudly; it is that they break quietly, an operator shrugs, retries with a different phrasing, and the maintainers never hear about it. Silent tool failures have been a live complaint on the Claude Code issue tracker for over a year. Self-reporting closes that gap without asking anyone to file anything.
The obvious question is what leaves your machine and when. Anthropic has not published the payload schema alongside the announcement. If you run Claude Code against regulated or client code, treat that as an open item to confirm before you assume the default is fine.
anthropic-sdk-python Hits v1.0 and Moves to httpx2
The Python SDK reached v1.0.0 on August 20, and it is a real major version, not a marketing one. The HTTP layer moved from httpx — no longer actively maintained — to httpx2, an API-compatible fork maintained by the Pydantic team. Same classes, same behaviour, security fixes included.
What breaks: Python 3.10 or later is now required. The legacy Text Completions API is gone. The temperature, top_p and top_k parameters are removed from Messages methods. The tool runner’s client-side compaction_control is removed. If you construct a custom http_client, Timeout or transport object, build it from httpx2 now.
One practical gotcha for anyone with observability wired in: tracing and mocking libraries that patch httpx by name will silently stop seeing your requests. Call httpx2.alias_httpx() at startup if that describes your stack. Pin your version deliberately — a anthropic>=0.x constraint in a requirements file will happily pull v1.0 into a Python 3.9 build and fail somewhere unhelpful.
Admin API User Management Goes GA, Mid-Conversation System Messages Land
Two platform changes worth wiring into your provisioning story. First, the Admin API user-management endpoints for Claude Enterprise organizations — members, invites, groups and custom roles — are now generally available. That is the piece that lets you drive Claude seats from your identity provider instead of a console, which is the difference between an onboarding checklist item and an automated one.
Second, mid-conversation system messages are live on Claude Fable 5, Mythos 5 and Opus 4.8, across the Claude API, Amazon Bedrock and Google Cloud. You can now change the operating instructions partway through a conversation without rebuilding the whole context. For long-running agents that switch modes — research, then draft, then review — this removes an entire category of prompt-stuffing workaround.
Thirty-Three Days Until Sonnet 4.5 Retires
The counter only moves one way. Claude Sonnet 4.5 retires September 29, 2026, and after that date calls against claude-sonnet-4-5 fail outright — the same hard stop that hit Sonnet 4 and Opus 4 on June 15 and Opus 4.1 on August 5. Opus 4.5 is unaffected.
If you are also doing the SDK v1.0 upgrade above, do both greps in one pass: the pinned model identifier and the removed sampling parameters live in the same call sites. Search repos, config stores, notebooks, CI secrets and vendor integrations — not just the services you believe are live.
Claude Academy Opens With 355 Free Resources and No Login Wall
Claude Academy went live on August 20 — a free official learning hub spanning roughly 355 tutorials, courses, cookbooks and use-case examples across Claude.ai, Cowork, Claude Code, Claude Tag and the Claude Platform. Courses are accessible without signing in, with material available in multiple languages including Japanese.
Beyond product training it carries broader AI-literacy content: how the models work, where they fail, and how to decide when to reach for one. There are role-based tracks for developers, creative professionals and leaders, plus agentic architecture guidance. Users with an account can track progress and earn completion badges that post to LinkedIn.
The no-login-wall decision is the notable one. Vendor training normally sits behind a form so the leads get captured. Publishing 355 resources openly reads as a bet that the constraint on enterprise adoption is competence, not awareness — and if you are rolling Claude out internally, it saves you writing your own onboarding curriculum.
MCP Has Quietly Become the Default Integration Layer
A useful marker on the protocol Anthropic donated to the commons. MCP now spans every major AI host, with mature multi-language SDKs and a growing tooling layer around it. Third-party registry counts have gotten large enough to be nearly meaningless — Glama indexes 71,000+ servers and the MCP Toplist tallies close to 101,000 across combined registries, though both figures include abandoned and duplicate listings. The more defensible number is roughly 5,800 actively community-published servers.
Host support is the part that matters. Claude Desktop was first and remains the reference implementation; Cursor and VS Code ship native MCP clients, Zed and JetBrains followed, and several agent frameworks speak it out of the box. Gateway products like MCP360 now front 100+ tools through a single endpoint.
Worth noting what this means competitively: Anthropic authored the standard that makes its own connectors portable to rivals. That was a real strategic concession, and it is why MCP won.
Backers Reportedly Targeting $2 Trillion for an October Listing
Anthropic investors are said to be aiming at a $2 trillion valuation for an October 2026 IPO, which would make it the largest public offering on record — ahead of SpaceX. The company is privately valued at $965 billion following its $65 billion Series H in May, and filed a confidential S-1 on June 1, with Goldman Sachs, JPMorgan and Morgan Stanley leading and a Nasdaq window reported.
Reported annualized revenue run rate has passed $65 billion, more than 7x end-of-2025 levels, with full-year 2026 projections circulating in the $100–120 billion range. Every one of those numbers is reporting, not audited disclosure. Anthropic has confirmed neither the timeline nor the valuation, and roughly doubling a private mark at listing is an ambition, not a price.
$45 Billion for Compute That Arrives in 2027 Is an IPO Document, Not a Capacity Plan
Read the Nscale deal on its own terms and the arithmetic is strange. Anthropic is committing roughly $45 billion over six years — comfortably more than its reported annualized run rate — for 460 megawatts that does not serve a single request until late 2027. Nothing about this deal helps a developer who hit a rate limit this morning. Nothing about it shortens the next incident.
So what is it for? Two things, and both are about the eight weeks between now and October.
The first is a supply story for investors. A company going public on a growth narrative has to answer the obvious bear question: what happens when demand outruns the chips you can get? A signed, multi-year, named-supplier commitment on frontier silicon is the cleanest available answer. It converts “we hope to scale” into a line item with a counterparty. Underwriters love that. It is far easier to defend in a roadshow than a slide about efficiency gains.
The second is denial. Vera Rubin capacity is finite in 2027 and every frontier lab wants it. Locking 460 megawatts is as much about the capacity OpenAI does not get as about the capacity Anthropic does. That is a real strategic move, and it is also why the number is large enough to be uncomfortable — you do not deter a rival with a modest booking.
The risk is straightforward and worth naming honestly. This is a fixed obligation against variable demand. If the run-rate trajectory holds, $45 billion over six years looks cheap in hindsight and everyone moves on. If inference costs keep falling, or model efficiency improves faster than usage grows, or the enterprise adoption curve flattens the way it has for every prior enterprise software wave, then this is a very large committed cost sitting on a public company’s balance sheet with public shareholders reading it quarterly. Private companies can absorb a bad capacity bet in silence. Public ones cannot.
The practical read for anyone buying Claude at scale: this is a company that has now priced in enormous growth and told the market so. That is mostly good news for you — it means aggressive capacity investment, and it means pricing pressure downward as utilization has to be filled. But it also means the commercial posture post-IPO will be shaped by quarterly numbers in a way it has not been until now. If you are negotiating a multi-year agreement, the window where terms are set by a private company optimizing for adoption rather than a public one optimizing for margin is measured in weeks, not years.