AI news, November 19, 2026: Anthropic IPO, Gemini 4 Neon, Grok Bot books trips, Cloudflare Clairic
Short answer: The main stories of November 18 to 19: Anthropic prices its IPO at 83 dollars a share and starts trading today, Google releases Gemini 4 Neon at 3 dollars per million input tokens, Grok Bot can book trips and Cloudflare makes Clairic open-weight. Rumors about a Telegram agent on Hermes and a Claude Code leak issue are unconfirmed.

Illustration by Claude (AI), drawn with code: a stock-exchange bell with a rising price (the IPO), a neon gemstone with a price tag (Gemini 4 Neon) and an airplane with a hotel key card (Grok Bot booking trips).
Anthropic prices its IPO at 83 dollars a share; trading begins todayConfidence: high for price and date: the official account and several press sources; medium for proceeds, share count and valuation: sources differ and the figures are not final.
Anthropic has priced its IPO at 83 dollars a share, above the 70 to 78 dollar range in the prospectus; trading on Nasdaq begins today, November 19 (Anthropic's account on X, 1.6 million views). Press reports and a news site citing TechCrunch, CNBC and Bloomberg give the ticker as ANT, 825 million shares offered according to the prospectus and proceeds of about 68.5 billion dollars, or 94.5 billion dollars with a 26 billion dollar private placement. The valuation comes to about 1.08 trillion dollars on shares outstanding or about 1.5 trillion dollars counting all options and reserves.
The number of shares sold has not been reported as final: @testingcatalog cites 980 million Class A shares, a different number from the prospectus. A listing is not product news, but the figures from the earlier registration filing are in our article on Anthropic's S-1/A.
What this means: the price and the date are confirmed by the company itself and by the press. Total proceeds and valuation are estimates based on share counts that may still differ. For anyone using Claude in a business nothing changes in the product today; as a listed company Anthropic will have to publish quarterly reports, which become a new source on revenue and costs. This is not investment advice.
Sources: @AnthropicAI over de prijsstelling van de beursgang op X · @testingcatalog over het aantal Class A-aandelen op X · Anthropic prijst beursgang op 83 dollar per aandeel (StockAnalysis, gelezen) · TechCrunch: Anthropic prijst beursgang op 83 dollar
Google releases Gemini 4 Neon: 3 dollars per million input tokens, 1 million contextConfidence: high for availability, price and context: Google's blog was read directly; medium for the benchmarks: figures from Google and outside indexes, not re-run.
Google announced Gemini 4 Neon on November 18 (774 thousand views for the launch post by @OfficialLoganK). The model is in the Gemini API and Google AI Studio (paid tier), in Google Antigravity and the Gemini app for Pro and Ultra subscribers and in Android Studio; access for cyber defenders through the Fairwind program follows over the next few days. Google's blog, which was read, lists 3 dollars per million input tokens, 0.15 dollars for cached input, 15 dollars per million output tokens and a context of 1 million tokens. At the high thinking level the cost per task is said to be 32 percent lower than Gemini 4 Carbon.
The performance figures come mostly from Google itself: 58.6 points on Artificial Analysis's Coding Agent Index (first place), 61 on its Intelligence Index (third place) and 65.2 percent on Terminal-Bench 4.1. @kimmonismus reports that Neon is 6 points above GPT-6.2 Astra and above Claude Fable 5.2 on that coding index (50 thousand views); that is a third party's benchmark and was not checked here. Cursor reports 63.7 percent on CursorBench 4.1 at the same price (117 thousand views).
What this means: price and availability are in the announcement. Whether Neon beats your current choice in your own work only your own tests can say: the benchmark figures come from the maker and from outside indexes that each measure different tasks. Earlier models in the same series are in our article on Gemini 4 Carbon and our article on Gemini 4 Hydrogen.
Sources: Google: Introducing Gemini 4 Neon (gelezen) · @OfficialLoganK over de lancering van Gemini 4 Neon op X · @kimmonismus over de Coding Agent Index op X
Grok Bot can book flights, hotels and rental cars; Grok 4.8 still in beta variantsConfidence: medium: own accounts and a quote from Musk; the article and booking details were not read; the fan quote is unconfirmed.
The account @bot reports that Grok Bot can now book flights, hotels and rental cars (260 thousand views). Elon Musk quotes an article by @poteto, Automate Anything with Grok Bot (1.18 million views, 17.8 thousand bookmarks), saying you should treat Grok Bot like a really smart employee (1.3 million views). The details of the booking (payment method, confirmation step) were not read; the description comes from the posts.
About the model itself Musk writes that Grok 4.8 is currently several different variants in beta and that it is not yet clear which will go to wide release (105 thousand views). A fan account adds that versions are quietly tested in Grok Bot; that quote comes from a video and is not confirmed. A timeline by @aapakari shows the expected release has slipped repeatedly between November 10 and 17.
What this means: once an agent can make payments or bookings, the approval rules matter more than the feature itself. Anyone who wants to connect this sets a spending limit and a confirmation step first. There is no release date for Grok 4.8. Earlier news about the build environment is in our article on Grok Build.
Sources: @bot over boeken met Grok Bot op X · @elonmusk over Grok Bot als slimme medewerker op X · @poteto over Automate Anything with Grok Bot op X · @elonmusk over Grok 4.8 in beta-varianten op X · @aapakari over de verschoven tijdlijn op X
Agent APIs: Perplexity Context and OpenAI background mode with a discount for waitingConfidence: medium: official Perplexity and OpenAI accounts; documentation, prices and terms were not read.
Perplexity introduces the Context API (160 thousand views): a list of web pages is converted into content for an agent in the format you want. According to @AravSrinivas Perplexity has released three APIs in a few days: Monitor for programmable alerts, Orchestrator for routing across models and Context for fetching pages (19 thousand views). Prices and terms of the Context API were not read. Teammates and Multiagent are now also in the iOS and Android apps for Max subscribers (38 thousand views).
OpenAI reports a background mode in the Responses API with a completion_window of 2 hours (30 percent cheaper) or 8 hours (40 percent cheaper) on GPT-6.2 Astra and Sol (137 thousand views). OpenAI also explains that the Codex sandbox has no network by default and that .git and .codex are write-protected (63 thousand views).
What this means: anyone running overnight jobs, such as summaries or briefings, can trade waiting time for a lower price; first work out how many tasks can really wait. The Context API is a possible alternative to your own scraping, but compare costs and terms first. Articles on the models are on our site: our article on GPT-6.2 Astra and Sol and our article on the Perplexity Orchestrator API.
Sources: @AravSrinivas over de Context API op X · @AravSrinivas over Monitor, Orchestrator en Context op X · @perplexity_ai over Teammates en Multiagent op mobiel op X · @OpenAIDevs over background mode op X · @OpenAIDevs over de Codex-sandbox op X
Cloudflare releases Clairic and Clairic-flash as open weights; Java4Shell calls for patchingConfidence: high for Clairic: Cloudflare's blog was read (the maker's internal figures); medium for Java4Shell and cf: only the official posts, the OpenJDK advisory was not read.
Cloudflare is making Clairic (27 billion parameters) and Clairic-flash (9 billion) open-weight, according to Cloudflare's blog, which was read. Both are extraction models that read images and PDFs (up to four images or pages per request) and are on Workers AI and Hugging Face; the license is not stated in the article. Until December 1 they cost 0.10 dollars per million input tokens; afterwards 0.24 dollars for Clairic and 0.10 dollars for Clairic-flash. The quality figures come from internal tests that Cloudflare itself calls directional.
@CloudflareDev reports a release candidate of cf, one command-line tool for the whole public Cloudflare API, with cf migrate for Wrangler projects (25 thousand views). @Cloudflare also reports Java4Shell, a critical memory-corruption flaw in OpenJDK's Base64 decoder that can lead to remote code execution, with six WAF rules for paid plans and a call to patch Java runtimes now (26 thousand views). The OpenJDK advisory itself was not read.
What this means: the 9B variant is small enough to try locally for document extraction; read the license first. Anyone running Java checks their own versions with java -version and looks up the patch version with the vendor. An earlier article on Workers AI is our article on the KV-cache leak in Workers AI.
Sources: Cloudflare: An updated Clairic and Clairic-flash are now open-weight (gelezen) · @CloudflareDev over cf, de CLI voor de Cloudflare-API, op X · @Cloudflare over Java4Shell en WAF-regels op X · @corywilkerson over Clairic op X
Qwen4-Max free for five days in Nous Portal, Hermes shows its backup model, Mistral Vibe on desktopConfidence: medium: official accounts and a developer post; no documentation or terms read.
Nous Research reports that Qwen4-Max, Alibaba's new flagship, is free for five days in Nous Portal, with native vision and a context of 1 million tokens (218 thousand views). According to another source the model stays API-only, without local weights. One account reports that Hermes in the canary build now shows when it switches to a backup model, with the real provider error attached, and that this arrives in the next release; the stable version stays at 0.23.2 for now.
Mistral reports that Mistral Vibe is now also available as a desktop app for macOS and Windows, running agents locally or remotely (167 thousand views); a price claim of 14.99 dollars a month comes from a secondary post and is not confirmed. Mistral earlier reported that Vibe and the Le Chonk model are free for three months in Europe. Google DeepMind's account reports Genie 4 in Project Genie with native audio and realtime editing; details are not verified.
What this means: free access is a chance to put a model next to your current choice with a fixed task, but the period is short. Check the terms for your data. Background is in our article on Qwen4-Max, our article on Hermes Agent 0.23.2 and our article on Mistral Large 4 Le Chonk.
Sources: @NousResearch over Qwen4-Max in Nous Portal op X · @BkashJosi over het backupmodel in Hermes op X · @MistralAI over Vibe op desktop op X · @MistralAI over Vibe en Le Chonk gratis in Europa op X · @GoogleDeepMind over Genie 4 op X
Unconfirmed: Telegram agent on Hermes, Claude Code leak issue, shared ChatGPT-Codex limit and paste-promptsConfidence: low: one or a few posts per claim; the GitHub issue was found but not reproduced; no primary confirmation.
The following claims circulate on X and are unconfirmed. @testingcatalog reports that Telegram is preparing its own AI agent, Dictator, said to run on Hermes internally (78 thousand views); Telegram and Nous have not confirmed this. A GitHub issue (number 111061) reports that in Claude Code 2.1.327 for VS Code prompts from other users appear; the issue exists, but @bebuu_ai reports it has not been reproduced and has been passed to Anthropic's HackerOne team (89 views). Smaller accounts say ChatGPT and Codex will share one usage limit from December; no official OpenAI post was found.
Warning: one account claims to have earned 7,350 dollars with a Grok Bot guide (119 thousand views), a pattern of hype with earnings claims. Another account shares a paste-prompt for Claude Code said to come from the team, and yet another a skill pack for bug bounties installed with two lines. A third describes Claude Code without a subscription through a gateway to another provider's free tier, which may break the terms. Do not install anything from such posts without reading the code.
What this means: do not adopt rumors, paste-prompts and install lists without a primary source. Anyone using Claude Code follows the issue and avoids sensitive prompts in the VS Code extension until there is a response. An earlier article on the previous version is our article on Claude Code 2.1.326.
Sources: @testingcatalog over Telegram Dictator op X · GitHub-issue 111061 over Claude Code 2.1.327 · @bebuu_ai over het issue en HackerOne op X · @Souradip3000 over een gedeelde limiet op X · @Bober_smart over verdienclaims met Grok Bot op X
Primary sources: Google: Introducing Gemini 4 Neon · Cloudflare: Clairic en Clairic-flash open-weight · Anthropic over de prijsstelling van de beursgang (X)


