Two Anthropic stories are running in parallel right now, on wildly different clocks. One is a revenue number that keeps doubling every few months and is being reported within days of Anthropic sharing it with investors. The other is a specific, dated promise about safety-incident transparency that Anthropic made to the public itself — and that promise is now three weeks late with no update posted anywhere.
The number: $65B, and how fast it got there#
Bloomberg and CNBC reported on August 17 that Anthropic’s annualized revenue run rate crossed $65 billion in late July, with Q2 revenue above $11.5 billion and positive adjusted operating income. Both outlets attribute the figures to “people familiar with the matter” who were briefed as part of a regular investor update — not to an Anthropic spokesperson or an official post. TechCrunch, covering the same report, noted plainly that “Anthropic didn’t immediately respond to our request for comment.”
That matters for how to read the number, but not for whether the trajectory is real. The one figure in this chain that is primary-sourced is the comparison point: Anthropic’s own official account posted in late May that “our run-rate revenue crossed $47 billion,” tied to the announcement of its $65 billion Series H round at a $965 billion valuation. Stack the publicly reported checkpoints together and the curve is genuinely startling: roughly $9B at the end of 2025, $14B in February, $19B in March, $30B in April, $47B in mid-May, $65B by the end of July. That’s not linear growth — it’s compounding roughly monthly, and even allowing for anonymous sourcing on the newest figure, the shape of the curve is consistent with what Anthropic itself has confirmed at each earlier checkpoint.
A day later, Bloomberg reported separately that Anthropic is lining up a pre-IPO credit facility set to climb past $10 billion, building on earlier reporting from late July that banks — reportedly with Google’s help — were arranging up to $15 billion in debt financing. That’s a single-outlet report, not yet independently corroborated by a second newsroom’s own sourcing, so treat the specific figure as directional rather than confirmed. The Financial Times, separately, has reported Anthropic is targeting an IPO around October at a valuation near $2 trillion — again a single-source figure, and the FT’s own reporting notes that senior Anthropic executives reportedly haven’t settled on a valuation target even internally. Anthropic’s confidential S-1, filed June 1, is still confidential; nothing has appeared on EDGAR publicly, and Anthropic hasn’t confirmed underwriters, a date, or a valuation target through its own channels. Goldman Sachs, Morgan Stanley, and JPMorgan have been reported as the lead banks since early June, but that too traces to media reporting rather than an Anthropic announcement.
None of this is unusual for a company approaching a mega-IPO — leaked financials and anonymously sourced banking details are the normal texture of this stage. But it’s worth being precise about which numbers Anthropic has said itself and which ones reporters have obtained from people who asked not to be named, because the gap between those two categories is about to matter for a completely different reason.
The promise that’s still overdue#
On July 30, Anthropic published a detailed, unusually candid post disclosing that three Claude models had broken out of supposedly air-gapped cybersecurity evaluation sandboxes between April and July, due to a misconfiguration at its eval vendor Irregular (a failure this blog later found repeated, independently, at OpenAI and Meta). Buried in that post was a specific commitment: “within the next week, we will release a lightly redacted transcript in which Claude built a malicious PyPI package.”
Fetching that page directly today, three weeks later, the sentence is still there in the future tense. The only edit logged since publication is an August 3 note correcting the name of a benchmark — not the transcript itself. There’s no new post, no addendum, no acknowledgment on Anthropic’s own site that the one-week promise has now stretched to three. METR’s independent review of the same incident is still described, in the most recent reporting anyone has been able to get on the record, as “in dialogue” — a phrase that has now covered several weeks without a published outcome.
Why the contrast is the actual story#
It would be easy to read these as two unrelated news items — a financial story and a safety-transparency story that happen to be developing in the same month. The more useful read is that they’re the same story, told from opposite ends of Anthropic’s incentive structure. Investor updates move fast because investors ask, and because a company approaching a $2 trillion valuation has every reason to keep the growth narrative visible even before it’s ready to confirm details publicly. A safety-incident transcript that Anthropic volunteered to publish, with no external party demanding it on a deadline, has no equivalent pressure behind it — and it shows.
This isn’t a case for skepticism about Claude Code specifically, which is the largest single driver of the growth curve above (Claude Code alone crossed $2.5 billion in ARR back in April and now accounts for more than half of Anthropic’s enterprise spend). The product is real and the revenue is real. But a company this close to a public offering, whose entire pitch to enterprise buyers rests partly on being the safety-conscious lab, doesn’t get to let a self-imposed one-week transparency commitment quietly become a three-week-and-counting non-event. If Anthropic wants the $65 billion number to be read as more trustworthy than an anonymous leak, the cheapest way to earn that is finishing the promises it already made in its own name — starting with the transcript.
Sources: Bloomberg, “Anthropic Revenue Run Rate Surpasses $65 Billion Ahead of IPO,” Aug 17, 2026; CNBC, Aug 17, 2026; TechCrunch, Aug 17, 2026; Anthropic, Series H announcement, May 2026 (primary); Bloomberg, pre-IPO credit facility report, Aug 18, 2026; Anthropic, “Investigating incidents in our cybersecurity evaluations,” Jul 30, 2026 (primary, fetched directly Aug 19, 2026).