Newsletter

Anthropic raises $81.3 billion in the largest IPO on record

The Claude maker sold 980 million shares at $83, above its marketed range, and begins trading on Nasdaq under the ticker ANT at a valuation near $1.5 trillion.

A dark broadcast studio with a wide curve of blue market-data screens below a large Nasdaq wordmark, showing index levels and price changes.
Lead image The Nasdaq MarketSite studio in Times Square, where the exchange hosts listing ceremonies. Anthropic's shares begin trading on Nasdaq under the ticker ANT. · Photo: Luca Marfè / Wikimedia Commons · CC BY 2.0
On this page

Anthropic priced its initial public offering at $83 a share on Wednesday, above the $70 to $78 range it had marketed. The company sold 980 million shares instead of the 825 million first planned, raising $81.3 billion. That tops SpaceX’s $75 billion and makes it the largest IPO on record by proceeds. At the offer price, the Claude maker is valued at about $1.5 trillion on a fully diluted basis. That is up from $965 billion when it raised $65 billion in May.

Trading on Nasdaq under the ticker ANT begins on Thursday. Chief executive Dario Amodei and other insiders sold nearly 3.4 million shares each, about $282 million apiece, according to the S-1 filing. The pricing came after a jittery stretch for AI stocks that followed the departure of Nvidia chief executive Jensen Huang. The debut now sets a reference point for OpenAI and other private AI companies.

Featured source: TechCrunch , SEC filings for Anthropic, PBC .

Nvidia delays its quarterly 10-Q while an internal investigation continues

Nvidia will not file its quarterly 10-Q alongside Tuesday’s results, finance chief and interim chief executive Colette Kress told analysts. She said only that the document would arrive “in the coming weeks.” The form carries legal-proceedings and risk-factor sections that the earnings release leaves out. UBS analyst Timothy Arcuri said management attributed the delay to the company’s internal investigation of Huang and to whether additional risk factors are needed. Nvidia also faces federal scrutiny over whether partners sold restricted AI chips to China.

Filed from: TechCrunch , Nvidia results .

Blue Origin will fly Google’s first heavy-lift Suncatcher mission

Google Research says a Blue Origin New Glenn rocket will place Project Suncatcher satellites in sun-synchronous orbit in the first quarter of 2029. Each satellite would carry roughly the computing capacity of a data center rack, built around Google’s TPU chips. The flight would be the project’s first dedicated heavy-lift launch. Google has already run prototype hardware in orbit. It describes the 2029 mission as a step in a long-term research roadmap rather than the start of a commercial service.

Filed from: The Verge , Google Research .

An audit of 2,686 banking apps finds serious flaws in more than half

Ostorlab examined Android, iOS and HarmonyOS apps from 487 banks in 64 countries, reading published builds rather than testing live accounts. It found at least one high or critical issue in 51.5% of the apps. Embedded secrets were a recurring problem: 18% of Android apps shipped a high-confidence key such as a cloud credential or private key. Only about 5% offered passkeys, and 12% carried fraud-detection code. Static analysis shows where protections are missing, not whether attackers have used them.

Filed from: TechRadar Pro , Ostorlab report .

Windows Server 2028 preview turns hardened settings into the default

Microsoft’s public preview of Windows Server 2028 applies a security baseline aligned with US Department of Defense STIG guidance during installation. The baseline replaces the usual post-setup hardening project. Cipher suites are restricted to modern options, key exchange uses hybrid post-quantum groups, Kerberos takes precedence over NTLM, and the native NVMe stack becomes default. Microsoft warns that the defaults can break legacy software. This build also has a known race condition that can crash the LSASS process during hybrid key exchange. Microsoft recommends disabling those groups until a fix arrives.

Filed from: Ars Technica , Microsoft Windows Server Insiders .

Cloudflare publishes open weights for two extraction models

Cloudflare released Clairic, a 27-billion-parameter extraction model rebuilt after its preview, and Clairic-flash, a 9-billion-parameter version, both free to download. The models return confidence scores, character spans and probabilities for each field, and now read PDF pages as well as photos. Cloudflare reports that image-reading errors on its internal receipt test fell from 207 to zero after it redesigned the decoder. The company says no public extraction benchmark exists yet and that its comparisons are self-reported. Both models cost $0.10 per million input tokens until December 1, when Clairic moves to $0.24.

Filed from: Cloudflare Blog , Model weights .

Internal chats show Meta staff flagged the same Muse privacy problems as users

Internal chats reviewed by Reuters show Meta employees warning that the Muse agent read encrypted messages users had not expected it to see. Staff also said it mixed context between accounts and sent unprompted replies. One staffer received two bank balances they did not recognise. Meta says the stray messages were hallucinations produced by instructions to keep engaging users between sessions, and that the flagged cases were isolated. Reuters said it could not establish whether the balances belonged to other people.

Filed from: Ars Technica , Reuters investigation .

From the Community

Google opens Genie 4 world model to AI Ultra subscribers

Google DeepMind is rolling Genie 4 into Project Genie for eligible AI Ultra subscribers. The model allows mid-session edits by prompt and adds audio that adapts as a scene changes. Worlds can now be published and remixed in a community gallery. Genie 3 gains sessions of up to three minutes, 360-degree panorama seeding and linked worlds. Google says generated output carries SynthID watermarking; fidelity claims remain vendor-reported.

Filed from: Google DeepMind .

A reverse engineer documents an agent-driven workflow

An engineer with two decades of hardware and software work published a walkthrough of LLM-assisted reverse engineering. The routine points an agent at a binary and has it record findings in Markdown as it goes. It decompiles through a Ghidra bridge, and adds debugger breakpoints or logging probes when disassembly stalls. The post describes one practitioner’s habits and tool choices rather than measured results.

Filed from: Slow Chronicle .

Continue reading

Complete index →