AI news, November 13, 2026: Anthropic IPO, Claude Meetings, Gemini 4 Hydrogen, Perplexity Decider

Anthropic starts its public offering: terms are out, trading from November 19 according to X postsConfidence: medium: Anthropic's post read on X; terms come from a financial news source and were not checked in the registration document; the trading date is a post on X

@AnthropicAI posted about the public offering ('Our initial public offering'), with a legal disclaimer about the registration with the US regulator SEC. According to our search round of November 13 the post had about 2.8 million views. TechCrunch reported on October 26 that Anthropic had formally filed for the offering.

According to Renaissance Capital (November 10), Anthropic plans to issue 825 million shares at 70 to 78 dollars per share, under the ticker ANT on Nasdaq, with pricing in the week of November 16. That is an underwritten offering of 61 billion dollars, next to a separate private placement of 26 billion dollars. At the midpoint of the range the fully diluted market value comes to about 1.2 trillion dollars. @testingcatalog writes that trading starts on November 19; this was not read back in an official document.

What this means: for users of Claude nothing changes technically because of this announcement. For organisations that rely on Anthropic for business processes it matters that the company will soon report publicly on figures and risks, so prices and terms will be tested more often by investors. This is not investment advice; the terms can still change before pricing.

Sources: @AnthropicAI op X (beursgang) · @testingcatalog op X (handel vanaf 19 november) · Renaissance Capital: noteringsvoorwaarden · TechCrunch: aanvraag beursgang (26 oktober)

Claude Meetings: Claude joins video calls in Google Meet as a visible participantConfidence: high: announcement read on X; details match two reports of November 10; Anthropic's product page was not read back

@claudeai announced Claude Meetings: Claude joins a video call visibly, answers questions in the meeting chat and works on documents, plans and whiteboards during the call. Afterwards notes, action items and the transcript are on the meeting page. According to our search round the post had about 4.2 million views.

According to reports by Gulf News and Axios of November 10, the feature works with Google Meet only for now, as a beta for Team and Enterprise plans. Claude announces itself in the meeting chat and, according to Gulf News, cannot join calls with external guests without admin approval; when an external participant joins, Claude leaves the call. Personal plans follow later.

What this means: an agent in a meeting hears everything that is said and can use connections to work apps. Decide beforehand who may access the notes and the transcript, and tell participants that Claude is listening. For connections to work apps see our overview of MCP servers and agent tooling. The same direction shows in Grok Bot 1.0, announced the day before for Slack, Teams and meetings.

Sources: @claudeai op X (Claude Meetings) · Gulf News: Claude Meetings · Axios: Claude Meetings (10 november)

Gemini 4 Hydrogen: Google model in Cursor and OpenClaw within hoursConfidence: low to medium: announcement and replies read on X; Google's documentation was not read back

@GoogleDeepMind posted 'Meet Gemini 4 Hydrogen' (about 1.1 million views according to our search round). According to @openclaw the model is usable in OpenClaw right away. A post about Cursor mentions 54.3 percent on CursorBench; that score comes from the post and was not recomputed. Google's documentation for this model was not read back.

What this means: a new fast model that is immediately available in widely used coding tools makes a comparison on your own tasks worthwhile. If costs are a concern you can test it next to your current model without switching harness. Treat benchmark figures from posts as an indication.

Sources: @GoogleDeepMind op X (Gemini 4 Hydrogen) · @openclaw op X (Hydrogen in OpenClaw)

Persistent agents at work: OpenAI describes internal management, a default limit of 8,192 tokens costs pointsConfidence: low to medium: posts read on X; OpenAI's piece and the measurement data were not read back

@kevins8 (OpenAI) shared a piece about managing 'fleets of claws', persistent agents such as OpenClaw. According to the post, agents are more like contractors than regular software because they can take action in the background on their own; that makes access and oversight a matter for enterprise IT. @openclaw points to it for OpenClaw Enterprise. OpenAI's full piece was not read back.

@joelniklaus reports that OpenClaw asks for at most 8,192 tokens per reply when the model configuration sets no limit. With DeepSeek-V4.1-Flash that cut replies short and, according to his measurement, cost 16 to 21 points on a harness leaderboard; the result is from a single measurement. @distortgeekin put Grok Bot 1.0 in a Slack team: on day one the bot reported 'done' nine times and six of those were wrong. A second bot was then added that only checks.

What this means: record the permissions of each agent and set an explicit output limit in every model configuration. A separate check on 'done' reports is easy to build; more background is in our overview of agentic AI. The figures from the two practical stories come from one user each and are not confirmed.

Sources: @kevins8 op X (agents beheren bij OpenAI) · @joelniklaus op X (limiet 8.192 tokens) · @distortgeekin op X (Grok Bot in Slack)

Perplexity opens Decider v1.3: small decision models for routing and security screeningConfidence: medium: announcement and replies read on X; model cards and license were not read back

@perplexitydevs released pplx-decider-v1.3 and pplx-decider-flash-v1.3, open-weight decision models (27B and 4B) with a new 'locate' function that extracts passages from text. They are available through the Decisions API, Hugging Face and OpenRouter, free for a week on OpenRouter. @alexatallah writes that with Decider Perplexity read 1 billion tokens from 7,000 agent sessions in one day and found 14 fixes for its own agent security rules; according to him both models have an Apache license. License and figures were not checked on a Perplexity page.

What this means: a small model can pre-screen agent session logs before a larger model or a person reads them. The free window makes a short test possible. Run such a test on a piece of log without sensitive data; context is in our overview of AI and cybersecurity.

Sources: @perplexitydevs op X (Decider v1.3) · @alexatallah op X (7.000 agentsessies) · @AravSrinivas op X (gratis op OpenRouter)

Unconfirmed: GPT-6.2, a coordinated outage report and a wave of near-identical postsConfidence: low: rumor and patterns in posts, no primary confirmation

After the OpenAI livestream 'Cyber at frontier', @kimmonismus wrote that he had hoped for GPT-6.2; it was not announced and remains a rumor. Several accounts posted almost the same text about a large Microsoft outage; the similarity points to a coordinated network and no official confirmation was found. Near-identical posts about Simular Box appeared at the same time; @steipete warned that marketers pay for positive posts about agents launching this week. This is not confirmation of paid promotion.

What this means: do not treat these posts as news. Check an outage on the provider's status page and wait for an official announcement from OpenAI.

Sources: @kimmonismus op X (GPT-6.2 niet gelanceerd) · @steipete op X (waarschuwing betaalde berichten)