How to read this radar
A roundup of what circulated on X and in official sources since 5pm yesterday (Brasília time). CONFIRMED means an official account or page; X SUMMARY means it only shows up in X posts and trend summaries; RUMOR means not confirmed yet. Yesterday's radar, with Gemini 4 Hydrogen and Anthropic's IPO page, is here.
5pm update: added the two Claude outages, ChatGPT Wallet, the Codex Ambassadors leads, OpenAI's robotics reshuffle, the Grok Imagine cuts, Deepnote's agent benchmark and Bland's Tofu-1.
Anthropic and Claude
CONFIRMED (5pm update): Claude had two incidents today, according to the official status page. The first, a login issue on claude.ai and Claude Code, ran from 6:58 to 8:58am Brasília time (status). The second, with errors on claude.ai, the API, Claude Code, connectors, Claude Meetings and Managed Agents, ran from 10:09am to 12:38pm; Anthropic warned that scheduled tasks and routines could run more than two hours late and some runs were skipped (status). If you run agents on Claude, check whether any morning routine was missed.
CONFIRMED: Anthropic published 10 recommendations on its engineering blog for designing, running and reporting model safety evaluations (Anthropic; post). It also streamed a live chat between Dario Amodei and JPMorgan's Jamie Dimon today, ahead of the IPO (Anthropic).
X SUMMARY: Anthropic started sending Claude Duet invites. According to TestingCatalog, each Duet user gets 3 invites, and people who join by invite get free Claude Pro time; TestingCatalog says a month, while the invite text going around says a week (TestingCatalog). Duet is still US only. A post from Japan says Duet works on the Team plan there, but we found no official confirmation (post). No expansion to Brazil so far.
RUMOR: Claude Fable 5.5 has supposedly been in internal testing for months, with partners already using it and some Fable 5.1 users being routed to it, and would land next week (post). Anthropic has not confirmed.
OpenAI
CONFIRMED (5pm update): OpenAI launched ChatGPT Wallet, which lets dots make purchases with approval on every transaction, for Pro subscribers in the US only. Becca already covered it in this MaxA post.
CONFIRMED: OpenAI introduced the regional leads of its Codex Ambassadors program, who run meetups and workshops in local communities (Pauline Narvas, reposted by OpenAIDevs).
PRESS: according to Fortune, OpenAI dissolved its standalone robotics team and moved the researchers to other groups; the company told Fortune it is not leaving robotics but speeding it up with its general models (post citing Fortune).
CONFIRMED: Tibo, from Codex, said the banked limits reset has been activated and should land in everyone's accounts soon (Tibo). In the same batch, ChatGPT's dots got faster, use less jargon and now follow custom instructions, including which actions need your go-ahead first, like "show me a draft before sending any emails" (ChatGPT).
CONFIRMED: after the Cyber at the Frontier keynote, Tibo summed up what OpenAI is building for security: automated patching and automated red teaming, with the recording available (Tibo).
X SUMMARY: Tibo asked how Codex could be improved "without saying model" and the question passed 10,000 replies, a good read on what developers want from the agent (Tibo).
RUMOR: GPT-6.2 Astra and Sol were reportedly pushed to next week for more tuning (post). The repeated posts claiming you can run codex --model gpt-6.2-astra are copy-paste chains with nothing official behind them (example). Treat them as spam.
Hermes, OpenClaw and agents
X SUMMARY: Deepnote released the Agent Initiative Benchmark, which drops 12 models as always-on agents into the Slack of a fictional company. The biggest failure was not knowing when to stay quiet: agents jump into conversations they don't need to and overuse tools. In the published ranking, GPT-6 Astra came first, GPT-6.1 Sol second and Claude Fable 5.2 third; code and data are open (Jakub Jurovych, Deepnote; X summary). Required reading for anyone running an agent fleet in Slack.
CONFIRMED: OpenClaw already runs Gemini 4 Hydrogen (OpenClaw). OpenClaw also shared a write-up by @kevins8 on the controls OpenAI uses to deploy persistent OpenClaw agents internally, a useful read for anyone running a fleet of agents (OpenClaw).
Grok and SpaceXAI
PRESS (5pm update): The Information reports that SpaceX's AI group laid off about 60 people from Grok Imagine, leaving a handful to maintain the product, with the focus shifting to coding tools and agents (The Information). There is no official statement. Grok 4.8 still has no confirmed date.
No new official announcement since 5pm yesterday. The Grok 4.8 soft rollout to Grok Bot users, expected this week, has not been confirmed yet, and the model now shows up in rumor lists for next week.
Voice
CONFIRMED (5pm update): Bland unveiled Tofu-1, a speech-to-speech model built for calls in noisy settings like highways, storms and crying kids in the background. It listens while it speaks and separates voices in real time; every sound in the launch video was generated by the model (Bland). Early access is by request (Bland).
CONFIRMED: Claude Meetings started rolling out in beta on Team and Enterprise plans (Claude; Anthropic). Claude joins the call as a visible participant, on Google Meet only for now, answers questions asked out loud or in chat (in writing by default), writes notes and action items and builds docs and whiteboards live with the team. It can use each person's connected tools with permission. During the beta, notes, transcripts and answers are free, and docs and whiteboards made in a call are free through February 10, 2027. It saves the transcript and summary, not the audio, and is not available to organizations with zero data retention, HIPAA-ready configuration or customer-managed keys (help center). On X, it is being read as direct pressure on Otter and Fireflies (X summary).
CONFIRMED: Fish Audio launched Fish Agent, a no-code platform to build voice agents that make and take calls: describe the agent, get a phone number or bring your own, pick the LLM and connect tools and webhooks. It runs on the full-duplex Fish Live 3 model, which handles interruptions, and the median time to go live is 32 seconds (Fish Audio).
CONFIRMED: Retell passed $100 million in revenue run rate, from $10 million to $100 million in 17 months, profitable and with only $4.6 million raised, according to Bing Wu, from the company (post).
Open opportunities
Bland Tofu-1 (early access): you can request access on the model page; Bland does not list country criteria or a deadline (request access).
Fish Agent (48-hour credits): for the launch, Fish Audio is giving away agent platform credits for 48 hours starting November 12 at 2pm Brasília time, to people who like, repost, bookmark and reply "Live 3" to the launch post (post). You can also claim a free voice agent through the official link (Fish Agent).
Murf Startup Incubator Program: $1,500 in Murf voice API credits for 3 months, for companies under 10 years old with fewer than 100 employees, a live website and a valid use case. No VC backing required; API resellers do not qualify. Answers come in 1 to 2 days (Murf).
Claude Meetings (beta): Team plan users can already bring Claude into Google Meet, since the feature is on by default; on Enterprise, an admin has to turn it on (help center).
Sources
x.com/testingcatalog/status/2121018480757797299; x.com/tk_researcher/status/2121116204781940851; x.com/JamesMalsawm/status/2120992533685981629; x.com/thsottiaux/status/2120999993079451919; x.com/ChatGPT/status/2120970988796170275; x.com/thsottiaux/status/2121036562058400200; x.com/thsottiaux/status/2121062974161092792; x.com/insanekrishnaa/status/2121144620201766950; x.com/dailytch/status/2121142148967907717; x.com/openclaw/status/2121098531205501401; x.com/openclaw/status/2120979006250213651; x.com/claudeai/status/2120973994191094179; claude.com/resources/articles/claude-meetings; support.claude.com/en/articles/18325154-use-claude-meetings; x.com/i/trending/2120976979784544609; x.com/FishAudio/status/2120924589429522457; x.com/_bingwu/status/2120925189647315177; fish.audio/agents; murf.ai/startup-program; status.claude.com/incidents/nn11wk485zfn; status.claude.com/incidents/34wq7gxnxvvm; www.anthropic.com/research/building-strong-safety-evaluations; x.com/AnthropicAI/status/2121317227618340990; x.com/AnthropicAI/status/2121272083863011714; x.com/paw_lean/status/2121188275490611334; x.com/NGLSHMUELI/status/2121302375696945564; x.com/theinformation/status/2121279457084244122; x.com/jakubjurovych/status/2121256499552915586; x.com/i/trending/2121270359353688214; x.com/usebland/status/2121302984802152941; www.bland.ai/speech-to-speech; maxassistant.app/en/news/releases/chatgpt-wallet-agent-payments-us/.
