AI for Intelligence

Anthropic tells investors its AI may soon speed up its own research

The signal

Anthropic says its AI may soon sharply speed up its own research

What happened. On 2 November Anthropic filed the first amendment to its IPO prospectus, with results to 30 September (Anthropic, 2026b). Revenue reached $17.0 billion in the third quarter, about 1,127% more than a year earlier, but growth from the second quarter slowed to 47%, after 144% in the quarter before (Capoot, 2026). The net loss for the quarter was $23.29 billion, and purchase and lease commitments, mostly for computing, total about $895.4 billion. A new risk factor says internal estimates for future models show it "may imminently meet" a threshold in its own safety policy: AI that sharply speeds up AI research (Bort, 2026). Anthropic says no current model is past that line, and that it is considering changes to its Responsible Scaling Policy, its rulebook for model risk, to keep humans in control of model development. We covered the first IPO filing on 26 October; the new facts are the third-quarter numbers and this safety warning.

Why it matters. A supplier now tells investors, under securities law, that its models may improve faster than its safety checks. If you build on Claude, expect faster model changes, and possibly new controls on how the newest models are used.

What to do. Ask whoever owns your Claude contract to read the risk-factor section of the filing and note what could change your terms or access.

Meituan's 1.6-trillion-parameter model costs $1.20 per million output tokens

What happened. Meituan, the Chinese delivery group, released LongCat-2.5 on 1 November (Meituan, 2026). It has 1.6 trillion parameters, with about 48 billion used for each word it produces, and it reads images as well as text. Meituan says it optimised training and inference on Chinese-made AI chips. The API costs $0.30 per million input tokens and $1.20 per million output tokens, a limited-time discount, and it works with agent tools such as Claude Code and Codex (Meituan, n.d.).

Why it matters. A model of this size, built on domestic chips and priced this low, keeps pressure on Western API prices. For European companies the low price comes with a question your compliance team must answer first: where your data is processed.

Claude's older models start carrying an invisible watermark today

What happened. From 3 November Anthropic adds its text watermark, a hidden pattern in word choice, to four older models: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 (Clark, 2026). Models launched from 2 August 2026 already carry it. The mark applies worldwide, in the API, the Claude apps and Claude Code, and survives copy and paste and some editing (Anthropic, n.d.). Only eligible organisations under EU law, such as regulators, media, researchers and companies with their own AI Act duties, can detect it today. We covered xAI's plan to watermark Grok text on 1 November; the new step is that Anthropic now marks models it released before August.

Why it matters. A regulator or newsroom can now show that text your team drafted with Claude came from a machine, even after light edits. Anthropic also warns that a mark does not prove who wrote the ideas, because proofreading and translation can leave one too.

Try this today

Translate your next client call live across 63 languages in Google Meet

What it is: Google Meet now detects who speaks which of 63 languages and translates each speaker into each listener's language in near real time, with matching captions (Google, 2026a).

Who it is for: Managers who run calls with foreign clients, suppliers or subsidiaries.

Access: Paid: Workspace Business Standard and Plus, Enterprise Standard and Plus, Frontline Plus, and Google AI Pro and Ultra. Guests from other companies can use it when an eligible user organises the meeting. Base plans allow 90 minutes per meeting.

First step: In your next call, click the translation control in the toolbar and mark the languages you already understand so Meet leaves them alone.

Caution: Organisations on the Scheduled Release track get it gradually from 2 November, over up to 15 days. Google cites regional and age limits without listing them, and translation does not work in recordings, dial-in calls or client-side encrypted meetings.

(Google, 2026a)

Turn meeting notes into an editable PDF or slide deck in one chat, for free

What it is: The Gemini app now creates, previews and edits documents, PDFs, slide decks and spreadsheets inside the chat, and opens them in Google Slides, Sheets or Docs (Google, 2026b).

Who it is for: Anyone who turns notes into client updates, board slides or trackers.

Access: Free: all Gemini app users with personal Google accounts, plus all Workspace customers, in every language the app supports.

First step: Upload one page of approved notes, ask for a one-page PDF summary, then highlight a paragraph and ask Gemini to turn it into a table.

Caution: Workspace admins have no switch for this feature, so agree which client files may go in. PDFs work on web and mobile, other formats on the web only.

(Google, 2026b)

Worth watching

Claude's new rules on tracking people take effect on 12 November

Anthropic's updated Usage Policy starts on 12 November (Anthropic, 2026a). It bans tracking people without their consent, including through analysis of data already collected, and bars Claude from deciding or recommending whom to investigate, arrest or charge. Journalism, legal research, content moderation and consented tracking such as fraud monitoring stay allowed. Investigations and OSINT teams should check their Claude workflows against the new text before that date.

OpenAI will merge ChatGPT's Chat and Work tabs, first for a small group

OpenAI told users by email that it will combine the Chat and Work tabs of ChatGPT into one screen, where you can move from a question to a finished presentation or website (Mehta, 2026). It starts with selected users and gives no date for everyone; Codex stays separate. Announced, not yet released. Yesterday's edition covered the 12% score of the best AI model on SemiAnalysis's expert test.

Repos worth a look

lexifina/global-law

global-law from Lexifina stores national law as a version history: each change is one law taking effect, on the date it took effect (Lexifina, 2026). Czechia, Spain, Switzerland, Georgia and Malta have their whole statute books loaded; other countries hold a few acts or none yet. A built-in server lets Claude or another AI assistant quote what a provision said on a given day, running on your own computer with no network calls. 1,720 stars on 3 November 2026. Forward it to your legal, compliance or regulatory-monitoring analyst. The repository itself is public domain (CC0), but each country's law text carries its own terms.

OpenBMB/MiniCPM-OCR

MiniCPM-OCR turns scans and photographed pages into text, including tables, formulas, charts and stamps, and a second model scores each region from 0 to 1 so you know which parts to check by hand (OpenBMB, 2026). The main model is small, 1.3 billion parameters, and there is a free API and a web demo. 781 stars on 3 November 2026, Apache 2.0 licence. Give it to whoever digitises archives, annual reports or patent files; it covers Chinese and English documents only.

The CERNY.AI read

The most useful AI document this week is a securities filing, and it was written for investors. Anthropic told them more about revenue, compute debt and model risk than any product page will tell a customer. That is the habit I want clients to build: read vendor filings, policies and help-centre changes as intelligence, because watermarks, usage rules and price cuts land there first. Teams in competitive and technology intelligence should give every critical AI supplier the same watch they give a competitor, with one named owner and a monthly note. Setting up that watch is part of my competitive intelligence consulting.

One question for your next meeting

If our main AI vendor changed its usage rules, its models or its watermark tomorrow, who on our team would notice first, and how many days would it take? The CI maturity quick check shows whether your supplier monitoring has an owner.

Sources

  • Anthropic. (2026a, October 8). 2026 Usage Policy update. https://www.anthropic.com/news/2026-usage-policy-update
  • Anthropic. (2026b, November 2). Form S-1/A registration statement, Amendment No. 1. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm
  • Anthropic. (n.d.). How Claude marks AI-generated content. Claude Help Center. https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
  • Bort, J. (2026, November 2). Anthropic warns that it's very close to RSI capability, aka self-building AI. TechCrunch. https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/
  • Capoot, A. (2026, November 2). Anthropic releases amended IPO prospectus with Q3 financials. CNBC. https://www.cnbc.com/2026/11/02/anthropic-ant-s-1-quarter-financials.html
  • Clark, P. (2026, November 1). Anthropic expands text watermarking to older Claude versions beginning November 3. Mashable SEA. https://sea.mashable.com/tech/56556/anthropic-expands-text-watermarking-to-older-claude-versions-beginning-november-3
  • Google. (2026a, October 22). Real-time speech translation in Google Meet expands to 63 languages with automatic language detection. Google Workspace Updates. https://workspaceupdates.googleblog.com/2026/10/real-time-speech-translation-in-google-Meet-expands-to-63-languages-with-automatic-language-detection.html
  • Google. (2026b, October 29). Create, preview, and edit files directly in the Gemini app. Google Workspace Updates. https://workspaceupdates.googleblog.com/2026/10/create-preview-and-edit-files-directly-in-the-Gemini-app.html
  • Lexifina. (2026). global-law [Computer software]. GitHub. https://github.com/lexifina/global-law
  • Meituan. (2026, November 1). Introducing LongCat-2.5. https://longcat.ai/blog/longcat-2.5
  • Meituan. (n.d.). LongCat-2.5 pricing. LongCat API Docs. https://longcat.ai/platform/docs/pricing/longcat-2.5
  • Mehta, I. (2026, November 2). OpenAI is finally merging ChatGPT's Chat and Work tabs. TechCrunch. https://techcrunch.com/2026/11/02/openai-is-finally-merging-chatgpts-chat-and-work-tabs/
  • OpenBMB. (2026). MiniCPM-OCR [Computer software]. GitHub. https://github.com/OpenBMB/MiniCPM-OCR