How to read this radar
This is the daily roundup of the main AI news circulating on X. We mark as CONFIRMED what came from an official company account or page, as X SUMMARY what only appears in X trend summaries (with no official confirmation we verified), and as RUMOR what has not been confirmed yet.
Google: Gemini 4 Carbon arrives, but not for everyone
CONFIRMED: Google announced Gemini 4 Carbon, which it calls its strongest model, focused on cybersecurity and long-horizon tasks; internally it ran for days on engineering problems (Google blog, Sundar Pichai). Per Google, Carbon ranks first on Artificial Analysis' Intelligence Index v5 at 59, one point ahead of the 58 for Opus 5.5 we covered in our story, and scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Access starts this week only for a subset of cyber defenders through the Fairwind Program; developers, businesses and consumers get it “soon”. Introductory price: $2 per million input tokens and $10 output, later rising to $4 and $20.
OpenAI: ChatGPT Pro 200 cut and a Codex teaser
CONFIRMED: people who had ChatGPT Pro 200 between September 22 and 29 keep the old allowance only through October 29; after that the plan moves to the lower allowance at the same US$200 per month (OpenAI help center). According to OpenAI's email quoted on Hacker News, ChatGPT Work and Codex usage drops from 20x to 10x Plus, GPT-6 Pro drops from 200 to 100 messages per week, and affected users get 62,500 credits (about US$2,500) valid until December 31 (Hacker News thread). X SUMMARY: Tibo, from the Codex team, said usage limits were reset again (post) and promised “good news tomorrow” (post). Nobody knows what it is yet. 5 p.m. update. CONFIRMED: you can now text your dot, ChatGPT's personal agent, over iMessage or RCS with text, photos or voice notes, and it will text you back proactively with updates (ChatGPT). You turn it on from the dot's profile under Texting (ChatGPT). According to Engadget, it is a preview for US Pro subscribers for now. RUMOR: Tibo's promised good news hasn't shown up yet, and X keeps betting on GPT-6.1 Astra; some even talk about a “GPT-6.1 Bel” (post). Nothing official.
Anthropic: a government discount, drug discovery and the Fable 5.5 that didn't ship
CONFIRMED: Anthropic signed a OneGov agreement with the US GSA giving federal, state and local governments 40% off Claude (Anthropic). It also published early results from partnerships with six drug developers and said it has started its own drug discovery work (Anthropic blog). X SUMMARY: this reportedly comes with the acquisition of Mithril Medicine, a one-person biotech run with Claude's help (X trend), and a study where analysts using Claude Code followed the AI's randomly assigned leanings in most choices without noticing (X trend). RUMOR: the so-called “Fable 5.5 day” came and went with no launch. What came out was a paper signed only by “Claude (Anthropic)” about an algorithm for the 3-SAT problem, and X connected the dots (X trend). On the IPO, the X summary highlights US$16.3 billion in first-half revenue (X trend). Claude Duet is still US-only, and our colleague already updated the full Duet story; the Artificial Analysis ranking also has its own story. 5 p.m. update. CONFIRMED: Anthropic's Frontier Red Team said an unreleased internal research model wrote full exploit chains on its own against Chrome, Safari, Android, Windows and Linux; in Chrome, 62 CVEs were patched as of version 156, all rated critical or high, including 13 sandbox escapes (Anthropic blog). Claude for Legal got a gallery of ready-made artifacts for contract review, cite checks and regulatory horizon scans (Claude). And Boris Cherny from Claude Code shared Utah's pilot: 34 developers across 10 agencies, 80% set up in under 30 minutes and 71% saving more than 20 hours over 4 weeks (post). X SUMMARY: Senator Elizabeth Warren sent a letter pressing on risks ahead of the IPO (X trend). RUMOR: a model called “Obsidian 3” showed up in Cognition's FrontierCode 2.0 data at 46.8% versus Opus 5.5's 45.5%, then vanished; X guesses Fable 5.5 or an Opus 5.6 (X trend), and some bet on GPT-6.1 Astra (post). RUMOR: X posts say US Max subscribers are coming off the Duet waitlist this week, with Pro next and no date outside the US (post). Nothing for Brazil yet.
GitHub: Copilot Fusion orchestrates several models for you
CONFIRMED: GitHub Universe brought Copilot Fusion, a router and orchestrator that picks the models for each task by itself: a single model, a cascade (one model drafts and a judge decides whether to escalate) or critique (several models review in parallel and the original revises) (GitHub docs). GitHub says Fusion Max came within two points of Claude Code with Opus 5.5 Max on Artificial Analysis' Coding Agent Index at about half the API cost, and paid plans get 30% off through November 30 (GitHub, GitHub blog). For anyone orchestrating agents, it is the same judge and reviewer idea, built into the product.
Agents: fewer swarms, one well-built agent
X SUMMARY: the guide from Anthropic's applied AI team keeps circulating on X, recommending you start with one agent, one context and one loop, and add subagents only for a measured reason such as parallelism or context overflow (X trend). Along the same lines, a paper called “Memory Tax” shows that sophisticated memory systems often make more mistakes than retrieving straight from the conversation history (X trend). For anyone orchestrating several agents, both messages agree: measure before adding complexity. 5 p.m. update. X SUMMARY: a Harvard study found coding agents raised lines of code by 30% and PRs by 23%, but review time rose 49% and completed tasks did not increase (X trend). At the other extreme, Axiom used dozens of agents in parallel to formalize the Classification of Finite Simple Groups in Lean, 43.1 million lines in 24 days (X trend). In short: swarms work when the result can be checked automatically.
Voice: NVIDIA ships open transcription
CONFIRMED: NVIDIA released Nemotron 3.5 Transcribe with open weights, a 30-billion-parameter MoE with about 4 billion active, transcribing 40 languages including Portuguese, separating up to 8 speakers and handling recordings up to 1 hour (Hugging Face, model, TechCrunch). Key detail for real-time voice builders: it is for recorded audio, not live conversation.
Hermes and OpenClaw
CONFIRMED: Nous Research's Hermes Desktop now generates sound effects with ElevenLabs right in the app, with one-click variations and a sound library (Nous Research). OpenClaw shipped patch v2026.10.8 with fixes for installs, stuck chats and missing replies (OpenClaw).
xAI: no Grok 4.8 yet
RUMOR: Grok 4.8 is still just a promise; xAI's docs still list Grok 4.7 as the latest, and the “next week” prediction made the rounds on X on October 24 (FrontierDot).
Open models
X SUMMARY: DeepSeek V4.1 Flash topped OpenRouter's weekly usage chart with 35.4 trillion tokens (X trend). Clover-35B, a diffusion language model, shipped with open weights under Apache 2.0 (X trend). RUMOR: a GLM-5.5-Flash listing showed up on OpenCode, but Zhipu has announced nothing (X trend).
Sources
OpenAI Help Center: About ChatGPT Pro tiers; Hacker News: Pro 200 usage change; Tibo (@thsottiaux) no X; Anthropic: OneGov / GSA; Anthropic: early progress in drug development; Nous Research: Hermes Desktop + ElevenLabs; OpenClaw v2026.10.8; X: Fable 5.5 day / 3-SAT; Google: Introducing Gemini 4 Carbon; Anthropic Frontier Red Team: vulnerabilities with Claude; Claude: artifacts for legal work; Boris Cherny: Utah pilot; ChatGPT: texting your dot; Engadget: texting your dot; GitHub Docs: Copilot Fusion; GitHub Universe 2026; Hugging Face: Nemotron 3.5 Transcribe; TechCrunch: Nemotron 3.5 Transcribe; X: Obsidian 3 on FrontierCode.
