News

Releases & Models

AI Radar 11/16/2026: Claude Code's Code Mode rules the weekend, Cursor details the Fireworks breach and Zoom offers up to $150K for voice startups

5pm update: OpenAI opens applications on Nov 23 for free frontier model access for up to 200,000 scientists, Tencent's Hy4 flash is free on Nous Portal for two weeks, OpenClaw 2026.11.4 fixes WhatsApp and Signal and Duet promises more countries "very soon", with no date. This morning: Claude Code's Code Mode, Cursor's breach report and Zoom AI Launchpad with up to $150K in credits. Fable 5.5, Grok 4.8 and GPT-6.2 remain rumors.

Tela de radar escura com sinais luminosos representando notícias de IA

How to read this radar

A roundup of what went around on X and official sources over the weekend, since Friday at 5pm (Brasília time). CONFIRMED means an official account or page; X ROUNDUP means it only shows up in posts and X trend summaries; RUMOR means not confirmed yet. Friday's radar is here. 5pm update (Brasília time): OpenAI opened free frontier model access for up to 200,000 scientists, Tencent's Hy4 flash went free on Nous Portal for two weeks, OpenClaw shipped 2026.11.4 and Duet's head of product talked about more countries "very soon".

Anthropic and Claude

CONFIRMED: Claude Code got Code Mode in version 2.1.323 (Nov 12), an experimental opt-in where Claude writes JavaScript to call several tools at once, in parallel, instead of one by one. Turn it on with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1, terminal only; it does not work with -p, the Agent SDK, the desktop app or cloud sessions (docs; changelog).

X ROUNDUP: Code Mode was the weekend's dev topic, with reports of 37% fewer tokens and 1.6x longer sessions under the same limits; those numbers come from posts and the X summary, not Anthropic (X summary). For Projects and the proactive coordinator, Becca already covered it in this MaxA post.

X ROUNDUP: Fable 5.2 reportedly scored 41.25% on Scale's Remote Labor Index, nearly double Opus 5.5, and using Fable as the boss of cheaper Opus subagents became a trend (X summary). If you are migrating, read the MaxA post on the Fable 5.2 400 error first.

X ROUNDUP: according to a site tracking the product, on Nov 14 everyone in the US on the Claude Duet waitlist (Pro, Max and Ultra) was let in at once; outside the US it is still closed (Pasquale Pillitteri). One user says Duet showed up on his Team plan, but the help page still says Team and Enterprise are not included (post). No Brazil yet.

PRESS (5pm): according to Innopark IT, Karan Ahuja, head of product for Claude Duet, said Anthropic aims to open Duet to more countries outside the US "very soon", with no date; new users get extra limit resets (three on Pro, one on Max and Ultra) that expire on Dec 15 (Innopark IT). For now it is still US only, no Brazil.

X SUMMARY (5pm): mathematician David Bessis saw his own name and email show up in Claude's reasoning during an incognito chat. The help page says those chats are not saved to history or used for training, but they are kept for 30 days and still use profile data (X summary).

OpenAI and Codex

CONFIRMED: OpenAI launched a science blog with physicist Nathaniel Craig explaining how he used GPT-6 Astra to propose a lattice definition of the Standard Model (OpenAI). Becca broke it down in this MaxA post.

CONFIRMED: Tibo, from the Codex team, said early today that the team "might have actually fixed" the "model at capacity" errors that had been blocking users (Tibo). He also asked for feedback on Daybreak, OpenAI's cybersecurity package with models, Codex plugins and red/blue team access (Tibo).

RUMOR: posts on X claim GPT-6.1 Astra was held back over safety and alignment concerns, and some say the best models are staying inside the labs; none of this was confirmed by OpenAI (post; post).

CONFIRMED (5pm): OpenAI is expanding ChatGPT for Academic Researchers to up to 200,000 scientists worldwide, with free access to frontier models, training and support; applications open Nov 23 and, by default, program data is not used for training (OpenAI; program page). In a study of more than 1,800 participants, scientists reported nearly doubling their productivity, according to OpenAI itself (OpenAI).

CONFIRMED (5pm): OpenAI is giving Apollo Research, METR, Transluce and Goodfire deep access to internal models and systems to audit alignment and safeguards, which it calls embedded assessments (OpenAI). And ChatGPT for Healthcare got more capable computer-use agents and 20+ healthcare plugins, used by Abridge, Cigna and CVS for prior authorizations, referrals and discharges (OpenAI).

X SUMMARY (5pm): Tibo's chart shows Codex "model at capacity" errors dropping from 6.82% of daily users in mid-October to 0.085% (X summary).

Cursor

CONFIRMED: on Nov 13 Cursor published the full report on its security incident. It started with a key Cursor had issued to Fireworks that was exposed on Fireworks' side; with it, the attacker copied old agent traces (from sessions without Privacy Mode), found unredacted credentials and reached internal tools. On Oct 29 they published fake copies of the Claude Code extension (versions 2.1.312 and 2.1.313) on the cache of Cursor's marketplace, live for about 11 hours and potentially reaching up to 5,000 users. Cursor says fewer than 0.01% of customers had API credentials exposed, the incident has been contained since Nov 1 and it moved all inference off Fireworks (Cursor; post). Anyone who installed those versions should follow the notice's steps, since uninstalling alone may not be enough.

Grok and SpaceXAI

X ROUNDUP: Grok Bot reportedly passed 800,000 paid weekly users, double since its August launch, and an Elon demo of Grok Bot turning a 23-hour Starship livestream into a highlight reel went viral (X summary). Lauren Tan of SpaceXSI also made the rounds explaining her team shipped 11,000 pull requests in a month with only 0.5% reverted, built on lint, tests and verification before scaling agents (X summary).

The Grok invisible watermark on/off story already has its own MaxA post by Becca. Grok 4.8 still has no confirmed date.

Hermes and agents

CONFIRMED: Teknium of Nous Research showed how to turn off Hermes Agent's automatic self-improvement review while keeping memory: hermes config set auxiliary.background_review.enabled false (Teknium). Handy for anyone running a fleet of Hermes agents who wants to cut background calls.

X ROUNDUP: Teknium also advised swapping GPT-6 Luna for DeepSeek Flash or GPT-6.1 Terra in Hermes, saying Luna thinks more, fails more and costs the same (Teknium).

OPPORTUNITY (5pm): Hy4 flash, Tencent Hunyuan's new model for documents, slides and computer use, is free on Nous Portal for two weeks (Nous Research; Nous Portal). After that, Tencent lists $0.07 per million input tokens and $0.281 output, and says it matches Claude Haiku 5.5 in blind tests (Tencent Hunyuan).

CONFIRMED (5pm): OpenClaw 2026.11.4 fixes Signal messages skipped on reconnect, WhatsApp replies lost on settings reloads and damage to backups and memory merges, with 125 PRs from 90 contributors (OpenClaw; release notes). Anyone running agents on WhatsApp should update.

Voice

OPPORTUNITY: on Nov 10 Zoom opened AI Launchpad, giving up to $150,000 in Zoom AI API credits (transcription, translation, summarization, Decisions and voice agents) for 12 months, with no equity. There are three tiers ($5k, $50k and $150k), startups worldwide can apply, funding is not required and applications are reviewed on a rolling basis (Zoom; apply). Note: once credits run out, extra usage is billed to the card on file.

X ROUNDUP: devs keep playing with the Grok voice agent API; one hooked it to a Go2 robot dog that follows spoken commands (post).

Still in the oven

RUMOR: Claude Fable 5.5, Grok 4.8 and GPT-6.2 remain on the "still pending in 2026" lists, with no official date from any of the three companies (post; post). Treat it as speculation until there is an announcement.

Sources

code.claude.com/docs/en/code-mode; code.claude.com/docs/en/changelog; x.com/i/trending/2121592703955730662; x.com/i/trending/2122270286309691851; pasqualepillitteri.it/en/news/33769/claude-duet-invites-three-per-user-waitlist; x.com/davicorn/status/2121435280402211091; x.com/OpenAI/status/2122123437141553630; x.com/thsottiaux/status/2122206990101536852; x.com/thsottiaux/status/2121845965997101118; x.com/Airo_07/status/2122171877250777175; x.com/lu_scogna/status/2121479064703676723; cursor.com/blog/security-incident-update; x.com/cursor_ai/status/2121411147714970092; x.com/i/trending/2122263070789480600; x.com/i/trending/2122121154781688307; x.com/Teknium/status/2122164612187804034; x.com/Teknium/status/2121881702247387455; news.zoom.com/zoom-ai-launchpad/; zoom.ai/ai-launchpad; x.com/paz_al3x/status/2121427564757451263; x.com/adityaa_behera/status/2121939310350061695; x.com/Harshit_io777/status/2121924460597133675; maxassistant.app/en/news/tools/claude-code-projects-proactive-coordinator/; maxassistant.app/en/news/tools/fable-5-2-four-changes-break-existing-integrations/; maxassistant.app/en/news/releases/confirmed-openai-gpt-6-astra-standard-model-physics/; maxassistant.app/en/news/regulation/grok-invisible-text-watermark-flip-flop-ai/; www.innoparkit.com/anthropic-claude-duet-waitlist-us-pro-max-ultra-november-16-2026; x.com/i/trending/2122377859314380931; x.com/OpenAI/status/2122377585523126615; openai.com/index/expanding-chatgpt-for-academic-researchers; x.com/OpenAI/status/2122380010287960095; x.com/OpenAI/status/2122359930628063248; x.com/OpenAI/status/2122419941756531150; x.com/i/trending/2122261408637206608; x.com/NousResearch/status/2122411675412463689; portal.nousresearch.com; x.com/TencentHunyuan/status/2122217770314301469; x.com/openclaw/status/2122379635434607039; docs.openclaw.ai/releases/2026.11.4.