News

Releases & Models

AI Radar 11/09/2026: Claude Fable 5.2 lands, Codex Pro with no 5-hour limit and Hermes Cloud in the browser

5pm update: Anthropic released Claude Fable 5.2 and Mythos 5.2 (already in Cursor), ChatGPT dots can now use several computers, Hermes Cloud opens in the browser and the Grok API adds a default watermark. Earlier: Codex Pro with no 5-hour limit, Cursor voice chat on Grok Voice and LiveKit's voice-startup program. Fable 5.5, GPT-6.2 and Grok 4.8 remain rumor.

Radar escuro com anéis ciano e feixe de varredura

How to read this radar

A daily roundup of what circulated on X and official sources, covering the weekend. CONFIRMED means it came from an official account or page; X SUMMARY means it only shows up in X trend summaries; RUMOR means it is still unconfirmed.

5pm update

CONFIRMED: Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 (announcement). Becca already covered pricing and access here: Claude Fable 5.2 arrives cheaper; Mythos keeps restricted access. Two extras: Boris Cherny said longer-tenured subscribers get more Fable limit resets (Boris), and Cursor already shipped Fable 5.2, which tops CursorBench at 60% on max effort (Cursor). Note this is Fable 5.2; Fable 5.5 is still RUMOR.

X SUMMARY: Apiiro researchers found over a thousand public .claude/settings.local.json files on GitHub holding real passwords and keys, saved when users approve a command with "don't ask again". The summary says Anthropic will add secret scanning (X summary). Anyone running agent fleets on Claude Code should check their repos.

CONFIRMED: ChatGPT dots can now connect several local computers and you pick which one handles each task, for example one that builds and tests while you keep coding on your laptop (OpenAIDevs). RUMOR: leak accounts say GPT-6.2 Astra testing is done and Sol is in final testing (post); OpenAI has announced nothing, and GPT-6.1 Astra is also still unreleased.

CONFIRMED: Hermes Cloud agents now open in a browser tab with the same interface as Hermes Desktop; you pick a name and model in Nous Portal and the agent runs 24/7 in its own container, with scheduled jobs (Nous; Nous Portal).

Grok and SpaceXAI: the Grok API now adds an invisible watermark by default since Nov 7, on models 4.20 through 4.7, ahead of the EU AI Act; admins can turn it off per model in the console (xAI docs; NewsLantern). Per press, Musk said on Nov 8 that xAI is working to make Grok Bot use fewer tokens and behave smarter (TokenPost). Grok 4.8 is still RUMOR.

X SUMMARY: Meta launched Muse Ember, a small model at $0.10 per million input tokens and $0.50 output, built for agent subtasks (X summary). Voice: Nuance Labs, founded by ex-Apple researchers, showed Emma, an avatar that reacts in real time to voice and video, with a public demo (X summary).

Duet: several users posted today that they got off the waitlist, pointing to a new invite wave (X chatter, example). The official page still says US only and Pro, Max and Ultra only (Anthropic help). No Brazil yet.

OpenAI

CONFIRMED: Codex lead Tibo clarified on Saturday that all Pro plans have no 5-hour limit and give 10x (Pro 100), 20x (Pro 200) and 50x (Pro 500, with ultrafast access) the usage of Plus (Tibo). He then asked whether a "Pro 50" would make sense, which is a poll, not an announcement (post).

CONFIRMED: OpenAI Developers added a Showcase tab to the OpenAI Developers plugin; eligible Pro and Business subscribers can claim API credits in billing settings (OpenAIDevs).

CONFIRMED: the OpenAIDevs "Tomorrow." teaser that had many expecting a new model pointed to an OpenAI x ModRetro Chromatic collaboration at the Supply Co. store on Nov 9, per press. It is not a model launch (post; News9live).

RUMOR: X users claim GPT-6.2 Astra and Sol are already served through quiet routing, based on knowledge-cutoff tests on Ultra Fast. That is user inference with no word from OpenAI (post).

Anthropic and Claude Code

X SUMMARY: developers complain Claude Opus 5.5 got "lazy" and burns quota faster since around Nov 3. Benchmarks look steady and Anthropic has not commented (X summary).

X SUMMARY: Claude Code cloud sessions now connect to the Claude Duet app so you can follow and approve changes from your phone (X summary).

Duet: no sign of expansion outside the US or general availability; still a research preview with a waitlist and a US account requirement (Centelha). RUMOR: Fable 5.5 and Opus 5.6 show up on X "models of the week" lists with no official announcement.

Cursor, Hermes, OpenClaw and agents

CONFIRMED: OpenClaw shipped v2026.11.2, a patch with no new features: it fixes stuck updates, missing scheduled reminders and duplicate chat widgets, with 113 PRs and 64 contributors (notes; post). For agent fleets, the vanishing-reminders fix is the one that matters.

CONFIRMED: Nous released the official Hermes Agent Security plugin, which maps what an attacker can reach, splits the code across parallel subagents and checks every finding against the code (Nous).

X SUMMARY: Nous also posted that using an agent that changes its models or code without telling you is "building on quicksand". It names no one, but the timing fits the OpenClaw model-swap debate (post). Read it as a jab, not a fact.

X SUMMARY: the orchestration debate continues: harness tests show the scaffolding around a model matters as much as the model, and people doubt delegating to subagents that lose context (harness; subagents).

Grok and SpaceXAI: Grok 4.8 still has no official release; Musk's last comment was Oct 30 ("taken longer than expected"). RUMOR: X lists put it this week, with no primary source (xAI notes).

Voice

CONFIRMED: Cursor launched realtime voice chat with agents on Grok Voice, in the IDE, web and iOS, with a privacy mode that keeps no data (changelog).

CONFIRMED: OpenAI added SIP to GPT-Live for inbound and outbound calls at $0.05 per voice-session minute, Retell now offers GPT-Live-1 and Grok Voice 3.0, and Google brought Gemini Live to Antigravity (hotwired.news, week 45).

Open opportunities

LiveKit Startup Program: up to $23K (one year of the Scale plan, $10K Cloud credits and $7K inference credits) for startups where voice, video or physical AI is core. It asks for under 50 employees, generally under $5M raised and pre-Series A, a live site, and production or launch within 6 months. No VC needed (livekit.com/startups).

OpenAI Developers: API credits for eligible Pro and Business subscribers via the Showcase tab (post).

Hermes Plugin Hackathon: until Nov 11, 5pm EST, $20K in prizes (announcement).

Careful: many "free credits" posts and resale offers of Claude Startup Program seats on X are scams or prohibited sales. Skip them.

Sources

x.com/thsottiaux; x.com/OpenAIDevs; docs.openclaw.ai; x.com/NousResearch; cursor.com/changelog/voice-chat; hotwired.news; livekit.com/startups; docs.x.ai; x.com/claudeai; maxassistant.app (Becca); x.com/bcherny; x.com/cursor_ai; x.com/OpenAIDevs (dots); x.com/NousResearch (Hermes Cloud); docs.x.ai (watermarking); tokenpost.com; support.claude.com.