The IPO Moved to November, and the Valuation Range Got Wider Instead of Narrower
Anthropic has pushed its planned listing from October to November, per Wall Street Journal reporting, while targeting a valuation of up to $2 trillion. The number that actually matters is the one nobody agrees on: outside estimates now run from $1.5 trillion to $4 trillion. When the spread is that wide this close to a listing, the market is not pricing a company — it is pricing a question.
The pressures named in the reporting are concrete rather than vibes. OpenAI has regained enterprise traction, open-weight models from Chinese developers and Meta are compressing per-token prices, and infrastructure costs keep climbing. There is also a slower-burn factor: the public debate over pacing AI development has become an investor-relations input, not just a policy one.
On the other side of the ledger sits the growth curve. Annualized revenue hit roughly $65 billion as of July, up from a $47 billion run rate in May and $10 billion for all of last year, and is projected to pass $120 billion by year end. The extra month buys a Q3 print to put in front of investors. If those numbers land, the delay reads as timing; if they wobble, it reads as something else.
A CCTV-Linked Account Published a 13-Revision Audit of Anthropic’s Privacy Policy
A social media account tied to China Central Television escalated its criticism of Anthropic yesterday, publishing a piece that walks through 13 privacy policy revisions since 2023. The specific claims: that the changes enable cross-border data transfers from users in Canada, Brazil, South Korea and the EU to the United States, that Anthropic expanded its listed data sources from three to six between June 2024 and September 2025, and that policy language permits sharing with US intelligence agencies when Anthropic deems it necessary.
Worth separating the verifiable from the framing. The revision count and the data-source expansion are documentable from Anthropic’s own published policy history. The characterization of intelligence sharing as happening “without legal procedures” is the contested part, and it is doing most of the rhetorical work in the piece.
The context is not subtle. In February Anthropic published a report alleging model distillation by Chinese companies and said it shared technical indicators with intelligence agencies. In June it told the US Senate that analysis of 28.8 million conversations and 25,000 accounts informed those findings. Whatever you make of the distillation claims, the disclosure that user conversations at that scale fed a government submission is the thread this campaign is pulling on — and it is a thread that exists.
Claude Leads About a Quarter of Anthropic’s R&D, and Anthropic Published the Oversight Numbers Too
Anthropic disclosed that the share of research and development work Claude “leads” — does in large part, under human supervision — went from none in February to roughly 26% in August. Six months, zero to a quarter. There were approximately 30,000 agents doing research and engineering work in August.
The oversight figures are the part worth reading closely, because Anthropic volunteered them. A real-time monitoring system checks every agent action before execution. Across more than one billion agent decisions in August, 0.002% were blocked — about one in every 47,000. That is either a very clean fleet or a very permissive filter, and the number alone cannot tell you which.
Anthropic tied the disclosure explicitly to recursive self-improvement, acknowledging that models accelerating their own development could make it “more challenging for humans to understand or control these systems”, and arguing that publishing the metric helps outsiders gauge how close frontier labs are to that threshold. Publishing a number you expect to go up, on a capability you are on record as worried about, is a defensible move. It is also now a benchmark other labs will be asked about.
Projects Is in Beta, and It Is a Coordinator Rather Than a Folder
The redesigned Claude Code Projects shipped to beta on September 17, and the shape of it is different from what “projects” usually means. You describe what needs to get done and Claude scopes the request, delegates the work, coordinates parallel threads, reviews the outputs and assembles the result. The unit of work is the project, not the conversation.
The implementation detail that makes it real: every thread is a full cloud session with its own branch and its own repository copy, and shared project memory and instructions reach every new thread automatically. That is the piece people have been hand-rolling with worktrees and copy-pasted context files for the better part of a year. Sessions keep running after you close the laptop.
Two caveats. Access started with a narrow slice of Pro and Max subscribers — those already using cloud sessions who have no existing projects on web or desktop — widening over the following week, with Team and Enterprise later. And because each thread is a full session, running several in parallel burns plan limits faster. Parallelism is not free; it is just spent differently.
Tool Search Turns a 77K-Token Startup Cost Into Under 9K
MCP Tool Search replaces eager tool loading with a lightweight search index plus on-demand fetch. Instead of dumping every tool definition from every connected server into context at session start, Claude Code queries the index and pulls only the definitions it needs. In Anthropic’s own measurement, a session with 50+ tools went from roughly 77K tokens before any work began to about 8.7K — an 85% reduction, preserving about 95% of the window.
The search side supports two modes: regex for precise matching when Claude roughly knows the tool name, and BM25 with semantic similarity for natural-language exploration. It is enabled by default on supported models and can be forced either way with ENABLE_TOOL_SEARCH. It disables itself in a few edge cases, notably a non-first-party ANTHROPIC_BASE_URL.
This is the least glamorous and most immediately useful change of the month. Tool bloat has been the quiet tax on every heavily-integrated setup — the more connectors you added, the less room your actual work had. If you trimmed your MCP config to save context, it is worth putting the servers back and measuring again.
Scoped API Keys, Agent Permission Policies, and an SDK That Drops Text Completions
Three platform changes worth a config review. First, the Console now issues personal keys and service account keys, giving you workspace-scoped access and per-account tracking. These keys carry the same permissions as the linked account and stop working when that account leaves the org — which quietly solves the offboarding problem where a departed engineer’s key kept billing for months.
Second, Managed Agents gained auto permission policies. The server evaluates each agent or MCP tool call and will run it, deny it, or pause for approval. Paired with ant beta:sessions connect in the CLI — which attaches your terminal to a live Managed Agents session so you can follow along, send messages, and approve or deny pending calls — the unattended-agent story finally has a human-in-the-loop that does not require building your own.
Third, the housekeeping: the SDK’s HTTP layer moved from httpx to httpx2, a maintained API-compatible fork, and v1.0 requires Python 3.10+ while removing the legacy Text Completions API and several deprecated parameters. Also new: computer use and browser use are now available on Google Cloud for Fable 5, Mythos 5 and Opus 5.
Smart Reports Reads 28 Days of Your Team’s Transcripts and Tells You What to Package
Smart Reports landed in beta for Claude Enterprise. It has Claude read up to 28 days of team transcripts and return them as workstreams, deliverables, cost, friction points, and candidate shared skills. During the beta, each org gets 10 free reports per month, resetting on the first of the calendar month.
That last output category is the interesting one. Usage analytics that tell you what a team spent are a commodity; a report that identifies repeated work worth packaging as a shared skill is a different product. It turns observability into a build queue, and it is a fair bet that the skills catalog is where Anthropic wants enterprise stickiness to live.
One constraint to check before you plan around it: Smart Reports is not available to organizations using customer-managed encryption keys, HIPAA configurations, or Access Transparency. Which is to say it is unavailable to a meaningful share of the regulated customers who would find team-level cost attribution most valuable.
MCP Adoption Went From Zero to 95% of Tool Calls in Seven Weeks
A community tracking effort measuring how tool calls route through client interfaces found a climb from 0% in late July to 95% by mid-September. Hugging Face Chat UI reached 100% — every tool call through their interface now goes over MCP. That is the kind of curve you get from a format winning on ergonomics rather than mandate.
The open source layer around it has settled into recognizable categories: servers that add capability (the official reference set, Context7, GitHub, Playwright), workflow and subagent frameworks (Superpowers, SuperClaude), and cost trackers like ccusage. The ClaudeWorld Public Works registry lists 13 active projects, 3 archived experiments and 1 upstream fork — small, but curated rather than scraped.
Pair this with the Tool Search story above and the shape of the year becomes clear: the protocol won the adoption fight first and is now dealing with the cost of having won. Everyone connected everything, context filled up, and the fix shipped. That is a healthier sequence than it sounds.
Two Disclosures, One Prospectus
Look at today’s top two official items side by side and notice they point in opposite directions. Anthropic is weeks from a listing at a target valuation of up to $2 trillion. It also just published a number showing that Claude leads a quarter of its own R&D, explicitly framed against the risk that models accelerating their own development make these systems harder for humans to understand or control.
That is not the disclosure profile of a company optimizing for a clean roadshow. A more conventional pre-IPO posture would describe the same fact as an efficiency story — look how much leverage we get per researcher — and never reach for the phrase “recursive self-improvement” at all. Anthropic reached for it, attached the oversight rate, and invited the comparison to other labs. Whatever else it is, it is consistent with the thing the company has claimed to be since it was founded.
Here is the uncomfortable part, though. The 0.002% block rate is presented as reassurance, and it might be. But a monitor that blocks one decision in 47,000 is either watching a fleet that almost never tries anything it shouldn’t, or it is a filter tuned to catch a narrow class of clearly-bad actions. Those are very different situations producing an identical statistic, and nothing in the disclosure lets an outsider distinguish them. The metric is a good-faith attempt at transparency that happens to be un-auditable from where the rest of us sit — which is the same structural problem as the threat intelligence report covering only the misuse that got caught.
Then there is the China story, which is the version of this argument conducted without any good faith at all. A state-media account is attacking Anthropic’s data practices precisely because Anthropic told a Senate committee it analyzed 28.8 million conversations in service of an allegation against Chinese firms. The underlying fact is real; the framing is adversarial; both things are true at once. And it lands on a company whose entire pitch to enterprise buyers is that it is the trustworthy one. Anthropic’s two most valuable assets right now are a growth curve and a reputation, and only one of them is in the prospectus. The valuation spread from $1.5T to $4T is, read generously, the market trying to price the second one.