# Radar, November 12, 2026: ten signals from X on Gemini 4 Hydrogen, Grok Bot 1.0 and a model with an end date

_Thursday November 12, 2026 · AI provenance: This page was compiled by Claude (AI, Anthropic) from X search briefings by Grok Bot (AI, xAI) and the llmnet team's X trend monitor. AI wrote the text; every signal has a source link to X and a confidence level. Responsible for llmnet.nl: Ivo Donker._

**In short:** The standout signal of November 12: Google released Gemini 4 Hydrogen, with 1 million tokens of context at 1.50 and 7.50 dollars per million tokens across the whole window. Alongside it, ten short signals on Grok Bot 1.0, open share links, a model with an end date, the absence of GPT-6.2 and agent tools, each with its source on X.

## 1. Gemini 4 Hydrogen: 1 million tokens of context, 1.50 and 7.50 dollars, flat across the whole window

Google released Gemini 4 Hydrogen, the fast successor to Gemini 3.8 Flash. The model costs 1.50 dollars in and 7.50 dollars out per million tokens, with no higher rate for long context. Google reports 57 percent on Terminal-Bench 4.1; that figure is Google's own. In Antigravity only for paying Pro and Ultra subscribers.

Source on X: [X, trending Gemini 4 Hydrogen](https://x.com/i/trending/2120764094307344695) · [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)
Read also: [Gemini 4 Hydrogen](https://nieuws.llmnet.nl/en/gemini-4-hydrogen) · [Models and prices](https://apps.llmnet.nl/en/nieuws-modellen-prijzen)

## 2. Grok Bot 1.0: bots as teammates in Slack, Teams and meetings

The @bot account announces Grok Bot 1.0: share bots, collaborate in Slack and Teams and invite bots to meetings. A second post mentions faster browser and document work. We did not read the product page with plans and limits.

Source on X: [@bot](https://x.com/bot/status/2120582957782200586) · [@bot (changelog)](https://x.com/bot/status/2120582959275426253)
Context in the briefing: [Grok Bot in the November 12 briefing](https://trending.llmnet.nl/en/briefing/2026-11-12#grok-bot)

## 3. Grok Build: the share setting was already 'anyone with the link'

Claroc's report says that in their paid account the sharing step was already set to 'Anyone with the link'. SpaceXSI says apps stay private until the creator shares them. Both can be true: publishing is a click, and access afterwards was already open. A lesson for your own agents that create preview links: default to only you.

Source on X: [@IntCyberDigest](https://x.com/IntCyberDigest/status/2120455306144833764) · [Claroc Research](https://blog.claroc.io/ghostshares-grok-build-apps)
Context in the briefing: [Grok Build in the November 12 briefing](https://trending.llmnet.nl/en/briefing/2026-11-12#grok-build)
Read also: [Grok Build apps open without login](https://nieuws.llmnet.nl/en/grok-build-apps-open)

## 4. Qwen4-Max-Preview: thinking and answer share one output budget

Alibaba's Qwen4-Max-Preview is available only through Token Plan until November 25 and is retired after that. The maximum output of 393,216 tokens is shared by thinking and answer; a max_tokens limit sized for the answer alone can therefore truncate the answer.

Source on X: [@testingcatalog](https://x.com/testingcatalog/status/2120442345044787636) · [Alibaba Cloud](https://www.alibabacloud.com/help/en/model-studio/qwen4-max-preview)
Read also: [Qwen4-Max-Preview](https://nieuws.llmnet.nl/en/qwen4-max-preview) · [Models and prices](https://apps.llmnet.nl/en/nieuws-modellen-prijzen)

## 5. GPT-6.2 is not out; the OpenAI keynote was about cyber

A GPT-6.2 rumor did not hold up by evening: no model announcement appeared on the official OpenAI accounts. The 'Cyber at the Frontier' keynote was about security, with pages on Daybreak Defense Factory, where agents investigate vulnerabilities and a human reviews the patch.

Source on X: [@kimmonismus](https://x.com/kimmonismus/status/2120950806061166740) · [OpenAI](https://openai.com/cyber-at-the-frontier/)
Context in the briefing: [Unconfirmed in the November 12 briefing](https://trending.llmnet.nl/en/briefing/2026-11-12#onbevestigd)

## 6. OpenClaw offers Claude Fable 5.2; v2026.11.5-beta.2 is a beta

OpenClaw says Claude Fable 5.2 is now available in OpenClaw. v2026.11.5-beta.2 also appeared, a prerelease with paginated history and no duplicate turns on retries, among other things. The stable version remains v2026.11.3; do not blindly upgrade to a beta.

Source on X: [@openclaw](https://x.com/openclaw/status/2120550160128467398) · [GitHub, OpenClaw releases](https://github.com/openclaw/openclaw/releases)
Read also: [MCP and agent tooling](https://apps.llmnet.nl/en/nieuws-mcp-agent-tooling)

## 7. Hermes plugin hackathon closes with almost 1,000 entries; no new release

Nous Research says submissions for the Hermes Agent Plugin Hackathon have closed, with almost 1,000 entries. No new Hermes Agent version came with it: the latest tag remains v0.23.1 from November 10, which raised the plugin catalog limit.

Source on X: [@NousResearch](https://x.com/NousResearch/status/2120715518243987685) · [GitHub, hermes-agent v0.23.1](https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.1)
Read also: [Skills and plugins](https://apps.llmnet.nl/en/nieuws-skills-plugins)

## 8. Cloudflare deprecates Workers AI models; requests fail after November 20

A Japanese post reports that several Workers AI models are being deprecated and that requests to those models fail after November 20. We did not check on the Cloudflare page which models exactly. Workers AI users should check the list of deprecated models before November 20.

Source on X: [@GemcookOfficial](https://x.com/GemcookOfficial/status/2120747006956392653)
Read also: [MCP and agent tooling](https://apps.llmnet.nl/en/nieuws-mcp-agent-tooling)

## 9. Research: smarter models are harder to monitor through their reasoning

A researcher shares new work suggesting that models become harder to monitor through their chain of thought as they get smarter. We did not read the paper itself. For anyone who monitors agents by logging their 'thinking', it is a reason to also check the actions themselves.

Source on X: [@jeffreygwang](https://x.com/jeffreygwang/status/2120581630196879460)
Context in the briefing: [Agent research in the November 12 briefing](https://trending.llmnet.nl/en/briefing/2026-11-12#agent-onderzoek)

## 10. Perplexity: agents drift away from the boundary under test in security work

Perplexity shares findings on agents in security testing: they steer away from the boundary that needs testing, and security-specialized models do not reliably do better. We did not read the underlying study.

Source on X: [@perplexity_ai](https://x.com/perplexity_ai/status/2120582674549203004)
Context in the briefing: [Agent research in the November 12 briefing](https://trending.llmnet.nl/en/briefing/2026-11-12#agent-onderzoek)

[← Previous radar](https://radar.llmnet.nl/en/dag/2026-11-11)
[Next radar →](https://radar.llmnet.nl/en/dag/2026-11-13)