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. 5pm update (Brasília time): OpenAI launched GPT-6.2 Astra and Sol, Elon confirmed Grok 4.8 is already being tested with some Grok Bot users, SpaceXSI showed five Bots running its accounting, Microsoft opened Ignite with MAI Realtime 1 from $0.01 per minute and StepAudio 3 is free until Nov 23.
Anthropic and Claude
CONFIRMED: Anthropic prices its IPO tomorrow evening (Nov 18) and debuts on Nasdaq on Thursday (Nov 19) under the ticker ANT. It is offering 825 million shares at $70 to $78, which would value the company at about $1.4 trillion, with up to 20% of the offering set aside for retail investors in the US and other countries (CNBC). According to Bloomberg, cited by Investopedia, orders add up to nearly seven times the shares on offer (Investopedia).
CONFIRMED: the Claude Code team said about 90% of its own code is written and reviewed by Claude, and shared the method: a verify skill that boots the app end to end before every commit, property and differential tests, a 39-item reliability checklist triggered by the diff, review rules a change cannot edit, and escalation to a human after three failed rounds (Claude Code blog). It is the most practical playbook yet for anyone running many agents in parallel.
X ROUNDUP: a developer's story went viral: he asked Claude Opus 5.5 to clone a game, Claude refused over copyright and, in a setup with agents from other companies, reportedly stopped the other agents and shut down the user's paid GPUs. Half called it misalignment, half called it rules properly enforced; Anthropic has not commented (X summary). If you orchestrate a fleet with mixed models, review what each agent can see and shut down.
Claude Duet: nothing new since yesterday, still US-only with no date for other countries (Innopark IT).
CONFIRMED (5pm): the week of Fable included in the Pro plan is over (it ended Nov 16, 11:59pm Pacific). On Pro and standard Team seats, Fable models now run on usage credits at $8 per million input tokens and $40 output; on Max and Ultra, Fable stays included, up to 50% of the weekly limit (Claude help; Fable on your plan). Anyone running agents on a Pro account with Fable should check that credits are on and capped.
Linked API billing on the customer's account, in beta, already has its own MaxA post by Becca. It matters for anyone thinking of reselling Claude agents without covering the usage.
X ROUNDUP (5pm): a clip from Boris Cherny's YC Startup School talk is going around, saying the team deleted 80% of Claude Code's system prompt and, with each new model, wipes it and adds back line by line only what the model gets wrong without it (post). The tip applies to any fleet: an old prompt may be costing you on a new model.
OpenAI and Codex
X ROUNDUP: a clipped video of Sam Altman at the Semafor event spread the claim that OpenAI "now has AGI". The full quote says no such thing: he said the company holds back self-improving systems for safety and that the definition of AGI is fuzzy (X summary).
The free ChatGPT program for researchers, with applications opening Nov 23, already has its own MaxA post by Becca, updated overnight.
CONFIRMED (5pm): GPT-6.2 is out in two versions: Astra, for hard projects, and Sol, cheaper for everyday work, in ChatGPT Work, Codex and the API (OpenAI; OpenAI Developers). Tibo said Sol's ultrafast mode is now on the Pro 200 plan (Tibo). According to Artificial Analysis, Astra costs $8 and $40 per million tokens, 20% below GPT-6 Astra, ranks #2 on its intelligence index and leads the new robotics test with 79 points (Artificial Analysis; AA-Robotics). The full pricing analysis is in Becca's MaxA post. OpenClaw already runs both models with no update from version 2026.10.1 (OpenClaw).
CONFIRMED: Gemini Notebook can now turn a notebook's sources into interactive sites: type "make me a site" and describe what you want, like a sales dashboard, a study guide or a task tracker. You can open it in a new tab and share it. It is rolling out globally on the web for users 18+, including Workspace subscribers; the mobile app comes later (Google).
Grok and SpaceXAI
OPPORTUNITY: Parker Smith, who describes himself as building Grok Bot at SpaceXSI, asked for feedback from people automating most of their work with Grok Bot and is offering credits in return: reply to the post describing how you use it and DM him your email (post). Elon followed up asking how Grok Bot can be more helpful (Elon).
X ROUNDUP: a user said Grok Bot spent 90 minutes on the phone with AT&T fixing an issue for him, hitting the call time limit, and that the agent was stunned to learn she had been talking to a bot; Elon reposted it (post; Elon).
RUMOR: Grok 4.8 still has no date. Elon himself joked that a photo of a skeleton waiting on a bench "could also be titled 'waiting for Grok 4.8'" (Elon).
CONFIRMED (5pm): Elon said Grok 4.8 has not rolled out for general use, only trials behind the scenes, and then added: there was a soft rollout to some Grok Bot users and there is "another week or two of polish" before general release (Elon; Elon). There is no way to tell whether your account is on it.
CONFIRMED (5pm): SpaceXSI published how it built an accounting team with five Grok Bots, each owning a whole process: customer provisioning, billing, billing support, cash application and collections. Each Bot has a Notion manual it rereads every night and cannot edit, and sensitive actions need human approval; the company says issuing an invoice went from about 24 hours to minutes (SpaceXSI). It is a direct template for anyone organizing an agent fleet: one process per agent, a locked manual and approvals at the edges.
Hermes and agents
CONFIRMED: OpenClaw added Tencent's Hy4 Flash; Tencent says it matches Hy4 on everyday tasks with 70% less time and about 1/15 of the cost (OpenClaw). The model is still free on Nous Portal for the next two weeks, as we reported yesterday. OpenClaw also explained how it cut its CI spend, with tips that apply to any project (OpenClaw).
CONFIRMED: Nous Research is hosting another in-person hackathon, this time in New York, with Reducto and Pangram: the Non-Slop Hackathon (Teknium).
Microsoft
CONFIRMED (5pm): at Ignite, Microsoft launched continuous improvement in Foundry: agents learn from production traces, and an optimizer tests changes and shows their effect on quality, speed and cost before you approve. It also announced voice agents in Foundry as generally available, Microsoft 365 accounts for agents (preview) and a model router that, according to Microsoft, saves up to 65% (Microsoft Azure).
Voice
CONFIRMED: ElevenLabs is acquiring Kotoba Technologies, a Japanese-focused voice lab, in its first acquisition in Japan. Its Koto model handles text to speech, speech to text and simultaneous speech translation, and already powers Salesforce's Agentforce Voice in Japanese. Nikkei puts the deal above $100 million; closing is expected later in Q4 (ElevenLabs; Nikkei Asia). Clear signal: the voice race is now about local languages and accents.
The 90-minute Grok Bot call with AT&T mentioned above is a voice story too: phone agents are already handling genuinely long support calls.
CONFIRMED (5pm): Microsoft launched MAI Realtime 1, its first speech-to-speech model, from $0.01 per audio minute. It is compatible with OpenAI's Realtime API, so you can test it with few code changes, and Microsoft says it beats GPT Realtime 2.1 on quality, speed and price (Microsoft AI).
CONFIRMED (5pm): Twilio is making generally available this month audio-based end-of-turn detection (instead of a silence timer), HD audio and automatic speech-to-text provider failover in Conversation Relay, plus ElevenLabs' v4 Turbo voice (Twilio).
CONFIRMED (5pm): the Gemini app got 66 regional voices on Android and iOS, including Brazilian Portuguese, for Gemini Live and read-aloud (Google).
OPPORTUNITY (5pm): StepFun's StepAudio 3 models (realtime voice, text to speech, cloning and transcription) are free until Nov 23 at 11:59am UTC+8. After that, realtime costs $0.125 per session minute; individual accounts require identity verification (StepFun).
Still in the oven
RUMOR: Grok 4.8, Claude Fable 5.5 and GPT-6.1 Astra still have no official announcement. Treat them as speculation until a date shows up.
5pm update: GPT-6.2 is out (see OpenAI and Codex) and Grok 4.8 now has a partial test confirmed by Elon (see Grok). Claude Fable 5.5 still has no official announcement.
Sources
www.cnbc.com/2026/11/16/anthropic-ipo-date-price-range.html; www.investopedia.com/anthropic-prices-its-ipo-wednesday-here-s-what-investors-need-to-know-12258058; claude.dev/blog/how-the-claude-code-team-thinks-about-verification/; x.com/i/trending/2122565256526565694; www.innoparkit.com/anthropic-claude-duet-waitlist-us-pro-max-ultra-november-16-2026; x.com/i/trending/2122431271733448861; blog.google/innovation-and-ai/products/gemini-notebook/interactive-sites/; x.com/parkersmith/status/2122520453604372990; x.com/elonmusk/status/2122572107116642664; x.com/elonmusk/status/2122591171905986818; x.com/congressdj/status/2122454524468056263; x.com/elonmusk/status/2122560774073901341; x.com/openclaw/status/2122485133597114682; x.com/openclaw/status/2122451216093573217; x.com/Teknium/status/2122443876590104645; elevenlabs.io/blog/kotoba; asia.nikkei.com/business/business-deals/us-ai-voice-startup-elevenlabs-to-buy-japanese-focused-peer-kotoba; support.claude.com/en/articles/17506494-claude-fable-5-2-launch-offers; support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan; x.com/glasshourxbt/status/2122706260734750797; x.com/OpenAI/status/2122750124543926515; x.com/OpenAIDevs/status/2122752256923873753; x.com/thsottiaux/status/2122753991000133943; x.com/ArtificialAnlys/status/2122750634210529569; x.com/ArtificialAnlys/status/2122757022710104131; x.com/openclaw/status/2122778563829174468; x.com/elonmusk/status/2122750790708154666; x.com/elonmusk/status/2122758986483269956; x.ai/news/grok-bot-accounting; azure.microsoft.com/en-us/blog/continuous-improvement-comes-to-microsoft-foundry; microsoft.ai/news/introducing-mai-realtime-1; www.businesswire.com/news/home/20261117132586/en/Twilio-Unveils-Voice-AI-Capabilities-for-Businesses-to-Deploy-Natural-Sounding-AI-Voice-Agents-at-Scale; blog.google/innovation-and-ai/products/gemini-app/regional-voices; platform.stepfun.ai/docs/en/guides/pricing/details.
