How to read this radar
A roundup of what went around on X and official sources since 5pm yesterday (Brasília time). CONFIRMED means an official account or page (or a press outlet with sourcing); X ROUNDUP means it only shows up in posts and X trend summaries; RUMOR means not confirmed yet. Yesterday's radar is here.
Updated at 5pm: added Telegram testing an agent that runs on Hermes, Qwen4-Max free on Nous Portal, OpenAI designing chips with AMD, Anthropic weighing a bigger IPO, Elon's timeline for Grok 4.8 and Microsoft's voice agents leaving preview.
Agents and orchestration
CONFIRMED: Cognition launched Devin Bots: persistent agents that own a whole goal, not just a task, and lead teams of Devins to deliver it. The launch example is a game studio with a single human employee and about 20 Bots, each owning one job (Unity, iOS, Android, testing, releases). Above them sits a lead Bot, the only one she talks to: four times a day it gathers the pending decisions into one Slack thread and she answers with emoji (Cognition). Bots come with the Max ($200/month) and Teams ($80/month plus seats) plans (pricing), and Fortune frames the launch as a response to Anthropic's Duet and OpenAI's Dots (Fortune).
X ROUNDUP: TokenWave launched SWE-CoWork, a test where an agent has to take over a teammate's work in progress from a short handoff. According to the summary circulating on X, no model fully completed more than 8% of tasks (TokenWave). The lesson for anyone running a fleet: the handoff is the weak spot, so invest in the context one agent passes to the next.
CONFIRMED: OpenClaw took Sign in with ChatGPT out of beta: you can use your ChatGPT plan, including Business, inside OpenClaw, already with GPT-6.2 Astra and Sol (OpenClaw). The project also said Microsoft's Autopilot is built on OpenClaw and is in preview for Frontier program customers (OpenClaw).
RUMOR (5pm update): Telegram is preparing its own agent called "Dictator", which shows up in the app's macOS betas and, according to what TestingCatalog found, runs internally on Nous Research's Hermes Agent, with support for multiple models, plugins and a dedicated virtual machine (TestingCatalog). Nous itself reposted the find (Nous Research), but neither Nous nor Telegram has confirmed a launch or a date.
OPPORTUNITY (5pm update): Qwen4-Max, Alibaba's new flagship model built for long-horizon work, with native vision and 1 million tokens of context, is free on Nous Portal for 5 days, in partnership with Inferact and the Qwen team (Nous Research; Nous Portal).
Anthropic and Claude
RUMOR: Anthropic sets its IPO price tonight and, according to the Wall Street Journal, it could top $80 a share, above the $70 to $78 range, on strong demand (Yahoo Finance, citing the WSJ). The company has not confirmed the final price; the Nasdaq debut is still set for tomorrow (Nov 19), as we reported yesterday.
RUMOR (5pm update): according to sources cited by CNBC, Anthropic is weighing upsizing the offering and setting the price well above $80. Demand reportedly reached about $400 billion from institutions and nearly $100 billion from retail, in non-binding orders. The final price comes after the market closes today and the shares debut tomorrow on Nasdaq under the ticker ANT (CNBC).
Claude Duet, checked again this afternoon: the official page still says US only, for Pro, Max and Ultra (Claude Help Center).
OpenAI and Codex
X ROUNDUP: after the GPT-6.2 launch, Sam Altman admitted the models are very smart but need more work to be "nice to work with", and said that will be a priority in upcoming releases (Sam Altman). Codex's Tibo explained why influencers and the public had such different takes: early testers got an older version, and the model changed a lot in the last 3 or 4 days (Tibo). He recommends Sol's ultrafast mode to Pro 200 and 500 users (Tibo).
CONFIRMED: the API usage dashboard gained an Analytics area with spend, rate limits, rate-limited requests and errors by model, project or key (OpenAI Developers).
Becca already covered on MaxA GPT-6 Cyber finding a serious Java flaw, ChatGPT for Sales and Astra for Law.
CONFIRMED (5pm update): OpenAI said that, working with AMD, its models found up to 14% improvement in power, performance and area on a flagship GPU design in under five months, building on a design AMD's engineers had spent years optimizing. The same approach helped improve Jalapeño, OpenAI's first custom inference chip (OpenAI).
CONFIRMED: OpenAI Developers explained the Codex sandbox: it runs commands in your environment but within filesystem and network limits you set, so a command that edits your repo doesn't automatically get internet access (OpenAI Developers). For anyone running a fleet of agents, that's the kind of guardrail that prevents damage.
X ROUNDUP: Tibo asked "how is 6.2 Sol treating you today?" and the post turned into a feedback drive with more than 2,000 replies (Tibo).
Free Gemini tasks in six countries, without Brazil, are in Becca's post.
5pm update: Gemini Live, camera and Nano Banana coming to low-memory Android phones are in Becca’s post.
Grok and SpaceXAI
X ROUNDUP: Elon kept pitching Grok Bot to companies: after accounting, he said the Bot can save millions by reviewing every expense, checking whether it is needed and finding better deals online or by calling suppliers directly (Elon).
X ROUNDUP (5pm update): Elon gave a timeline for Grok 4.8: there was a soft rollout to some Grok Bot users, but there is still "another week or two of polish" before general release (Elon).
Microsoft
CONFIRMED: besides the MAI Realtime 1 we covered yesterday, Microsoft launched the MAI model family at Ignite, including MAI Code 2 and Code 2 Flash, already available in GitHub Copilot (Microsoft AI; Microsoft Developer).
Voice
CONFIRMED: Interact introduced Interact-0, a full-duplex voice model that talks and listens at the same time and, according to the company, is the only one that knows when to speak in a group conversation without barging in. It has 31 billion parameters (3 billion active) and was trained on 9 million hours of audio (Interact). There is no public API yet, but it tackles exactly the problem of putting an agent into a meeting.
CONFIRMED: inclusionAI's (Ant Group) Ming-3.0-audio-flash ranked first on Artificial Analysis's speech reasoning test at 99.8%, ahead of Grok Voice 3.0, Gemini 3.8 Live and GPT-Realtime-2.1 (Ant Ling).
CONFIRMED: HeyGen launched Avatar 6, with gestures that follow the meaning of the script; it is already in the app and the API (HeyGen; docs).
OPPORTUNITY: Speko, which puts 31 voice providers behind one API, opened the first batch of its startup program: $1,000 in credits (about 185 hours of voice), a showcase spot and promotion. Idea-stage teams are welcome and no account or demo is needed to apply (Speko).
CONFIRMED (5pm update): voice agents in Microsoft Foundry Agent Service are now generally available. With them came OpenAI's GPT Live 1, which talks in full duplex and can hand tasks to other agents mid-conversation, and Microsoft Realtime Flash 1, in preview, focused on low cost for millions of calls; in Microsoft's internal testing it scored 70.9 versus 63.7 for GPT-Realtime-2.1 (Microsoft Foundry). Microsoft also made Contact Center Intelligence generally available, using MAI Realtime 1 (Microsoft).
Still in the oven
RUMOR: Grok 4.8 is still in a soft rollout to some users; per Elon, general release is a week or two away, with no official date. Claude Fable 5.5 still has no official announcement.
Sources
devin.ai/blog/introducing-devin-bots; devin.ai/pricing; fortune.com/2026/11/17/exclusive-cognition-bets-on-devin-bots-coding-agents-in-challenge-to-anthropics-duet-and-openais-dots/; x.com/TokenWaveai/status/2122713676511867009; x.com/openclaw/status/2122809472271958517; x.com/openclaw/status/2122877394331811874; finance.yahoo.com/markets/stocks/articles/anthropic-eyes-ipo-price-above-181809115.html; x.com/sama/status/2122798042734633452; x.com/thsottiaux/status/2122929634916679820; x.com/thsottiaux/status/2122929829863735594; x.com/OpenAIDevs/status/2122824860829176173; x.com/elonmusk/status/2122872199623414175; aka.ms/MAI-Ignite-2026; x.com/msdev/status/2122760423699268086; interact.inc/blog/interact-0/; x.com/AntLingAGI/status/2122972336303509613; www.prnewswire.com/news-releases/heygen-introduces-avatar-6-an-audio-driven-context-aware-ai-model-that-alleviates-repetitive-or-unnatural-motion-in-ai-generated-video-302942527.html; developers.heygen.com/docs/models/avatar-6; speko.ai/startups; x.com/testingcatalog/status/2123047003231858966; x.com/NousResearch/status/2123145750833995878; x.com/NousResearch/status/2123135537087181191; portal.nousresearch.com/; www.cnbc.com/2026/11/18/anthropic-ipo-ai-offering-price.html; x.com/OpenAI/status/2123100386638877054; x.com/OpenAIDevs/status/2123129469447880885; x.com/thsottiaux/status/2123113107048501273; x.com/elonmusk/status/2122758986483269956; techcommunity.microsoft.com/blog/azure-ai-foundry-blog/voice-agents-in-microsoft-foundry-agent-service-now-generally-available/4573013; www.microsoft.com/en-us/dynamics-365/blog/it-professional/2026/11/18/microsoft-contact-center-intelligence-now-generally-available-transform-customer-interactions/; support.claude.com/en/articles/17631495-get-started-with-claude-duet; maxassistant.app/en/news/tools/gemini-live-nano-banana-low-memory-phones/.
