Comic-style split image of a slow steam train beside a glowing bubble capsule racing past on the same track, illustrating Google's faster Gemini 4 Hydrogen model

Quick answer

Gemini 4 Hydrogen is Google’s new fast, lower-cost model in the Gemini 4 family, released on 12 November 2026 to replace Gemini 3.8 Flash. Google says it is 2.5 times faster than 3.8 Flash at its default setting, makes more than five times fewer fabricated claims, adds a fourth “xhigh” thinking level and costs $1.50 / $7.50 per million input/output tokens. WorldofAI’s own tests rank it just behind Claude Sonnet 5.5 and ahead of Gemini 4 Argon. The same week, Anthropic’s IPO founders’ letter said it may advance more slowly than it could for safety, Claude Meetings launched in beta, NVIDIA confirmed Jensen Huang’s resignation and OpenAI teased an unreleased cyber model.

Key takeaways

  • Gemini 4 Hydrogen is live now in the Gemini app, Google Search, AI Studio, the Gemini API and, for Pro and Ultra subscribers, Antigravity.
  • It is fast and capable, but not the cheapest. WorldofAI found it costs roughly a third of Sonnet 5.5 per task but about 2.5 times more than OpenAI’s cheaper option at the same setting.
  • Use the default “high” setting first. The creator found xhigh nearly doubles output tokens for a small gain.
  • Anthropic is not announcing a slowdown. Its founders’ letter says it is willing to move more slowly when safety requires it.
  • Claude Meetings lets Claude join Google Meet calls as a visible participant on Team and Enterprise plans (beta).
  • NVIDIA’s CEO change is confirmed by the company. The reasons reported in the press are allegations, and the separate Blackwell chip report is not linked to his departure.
  • OpenAI’s 77% cyber model is a teaser with no name, release date or system card.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

Some weeks bring one big launch. This one brought a new Google model, an unusual promise from Anthropic, a leadership shock at the world’s most important AI chip company and a glimpse of a very capable security model from OpenAI.

This round-up is based on WorldofAI’s 13 November news video. We focus on Gemini 4 Hydrogen, because it is the launch most people can use today, then cover the other stories with a clear line between what the companies have confirmed and what is still a claim. We checked the biggest stories against official sources, which we link to. Our earlier posts on Gemini 4 Carbon and the Gemini Krypton leak cover the rest of the family.

Watch: the source video

What this video covers: WorldofAI’s 13 November AI news round-up: the Gemini 4 Hydrogen launch with benchmarks, pricing and the creator’s own tests (Minecraft clone, landing pages, SVG and 3D scenes), Anthropic’s IPO roadshow and founders’ letter, Claude Meetings and Claude Managed Agents, NVIDIA’s CEO change and a forensic report on Blackwell GPUs, OpenAI’s unreleased cyber model and ChatGPT Dots update, a Fable 5.2 animated short, Seedance 2.5 Flash, Hark’s Handoff², Microsoft Ignite and Gemini on Apple CarPlay.

Video: WorldofAI (@intheworldofai), published 13 November 2026. Sponsored by Abacus AI (Abacus AI SuperComputer), which the creator mentions in the video. The creator’s benchmark results come from his own WoAI Bench tool. This summary and the opinions in it are our own; confirmed details are linked to official sources.

What is Gemini 4 Hydrogen?

Gemini 4 Hydrogen is the speed-focused member of Google’s Gemini 4 family, sitting below the frontier models Gemini 4 Carbon and Gemini 4 Argon. Google announced it on 12 November 2026 as a model built for “rapid execution” in everyday coding and agent work, and it replaces Gemini 3.8 Flash across Google’s products.

A few details stand out:

  • Four thinking levels: low, medium, high (the default) and a new xhigh, so you can trade speed and cost for deeper reasoning.
  • A 1 million token context window, shared with Carbon and Argon.
  • Built from Carbon: Google’s model card says Hydrogen is “adapted from Gemini 4 Carbon”. WorldofAI notes that earlier Flash models were described as based on the Pro model, and wonders whether this reflects a closer link between the two. That is the creator’s reading, not something Google has explained.
  • Designed to work in a team: in Google Antigravity, Carbon can plan while Hydrogen carries out coding and testing. WorldofAI reports Google’s claim that this setup reaches similar results using 24% less quota and finishing 19% faster.

If the leak stage passed you by, our earlier round-up covered Hydrogen when it was still a rumour.

How good is Gemini 4 Hydrogen, really?

On Google’s own numbers and on Artificial Analysis’s independent index, Hydrogen is a big step up from 3.8 Flash and roughly level with Gemini 4 Argon. Here are the main figures, with where each one comes from.

MeasureGemini 4 HydrogenComparisonSource
Speed vs Gemini 3.8 Flash2.5x faster (default setting)—Google
Artificial Analysis Intelligence Index51 at high, 54 at xhighGemini 4 Argon: 50Google, citing Artificial Analysis
Terminal-Bench 4.1 (coding)57%Gemini 3.8 Flash: 18.2%Google
Fabricated claimsMore than 5x fewervs Gemini 3.8 FlashGoogle
Hallucination rate8%Sonnet 5.5: 43%; OpenAI’s Sol: 52%Quoted in the video
Terminal-Bench 4.054%Sonnet 5.5: 64%; Sol: 56%Quoted in the video
Agent Arena debut5th place—Quoted in the video

The pattern is consistent: Hydrogen is strongest at speed, reading documents and images, and not making things up, while Claude Sonnet 5.5 still leads on hard terminal-based coding tasks. WorldofAI also says Hydrogen beats OpenAI’s GPT-6.1 Terra on most knowledge-work and agentic coding tests, and sits slightly behind Sonnet 5.5 in some areas.

⚠️ Warning: Benchmarks are a starting point, not a verdict. Vendors pick the tests they publish, and the hallucination figures quoted in the video depend on one evaluation. If accuracy matters for your work, run your own small test. Our guide How to Test an AI Model for Made-Up Answers Before You Rely on It shows you how.

What did WorldofAI’s own tests show?

WorldofAI ranks Hydrogen 7th overall on his own benchmark, ahead of Gemini 4 Argon and just behind Claude Sonnet 5.5, mainly held back by coding. His hands-on demos, which are his results rather than independent tests, included:

  • A Minecraft-style game with torches, a day-night cycle, block breaking and crafting, built in a single run of about 8 minutes.
  • Landing pages with scroll effects and varied typography, where he felt Hydrogen showed better design taste than OpenAI’s cheaper model.
  • An Xbox controller drawn as an SVG: Argon took about 6 minutes and roughly 83 cents; Hydrogen took about 4 minutes and roughly 42 cents, with what he judged a better result.
  • A 3D island scene with a sailboat, water and a small house, built in Antigravity using its swarm feature.
  • A presentation made from a YouTube video, which is the kind of everyday “agentic” work he thinks suits Hydrogen best.

His conclusion is balanced: Hydrogen is fast, reliable and strong visually, but not the cheapest model, and Opus 5.5 still adds small details that Hydrogen misses. He would choose Hydrogen as a daily driver for documents, presentations and data analysis rather than serious coding.

💡 Tip: The creator found Hydrogen works best on the default high setting, where it gives the most for the fewest tokens. Switching to xhigh nearly doubled output tokens for a modest gain. Our guide to choosing an AI reasoning effort level explains how to make that call for your own tasks.

How much does Gemini 4 Hydrogen cost and where can you use it?

Through the API, Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across the whole 1 million token window. According to Google, that is the same price as Gemini 3.8 Flash, so existing Flash users get the upgrade without a price rise.

WorldofAI puts it in context: Hydrogen costs roughly 32% of what Sonnet 5.5 costs per task, but about 2.5 times more than OpenAI’s cheaper model at the same reasoning level. Our guide to comparing AI models on token cost helps you work out what that means for your own workload.

Where you can use it

  • Everyone: Google Search, and the Gemini app (Google says free users get it through automatic model selection when it suits the prompt).
  • Developers: Google AI Studio and the Gemini API, on free and paid tiers.
  • Google AI Pro and Ultra subscribers: Google Antigravity, and Gemini in Google Sheets.
  • Businesses: Gemini for business, including Sheets for eligible subscriptions.

Rollouts vary by country and plan, so check the model picker in your own account before you plan around it. Google also says Gemini answers in AI Mode now show richer interactive visuals, which WorldofAI compares to ChatGPT’s interactive answers.

Is Anthropic really slowing down?

No. Anthropic has not announced a slowdown. Its founders’ letter says it is willing to advance more slowly than it could when safety requires it, even at a competitive disadvantage. That letter is part of the company’s IPO paperwork, and you can read it in full.

The letter gives a real example: Anthropic says it held back its Claude Mythos Preview model because of its cybersecurity abilities, and only released those capabilities (as Claude Fable 5) after building extra safeguards, giving up revenue and ground to competitors. It also commits to keeping Claude free of advertising.

WorldofAI is sceptical. He points out that the last time Dario Amodei spoke about slowing down, Anthropic went on to release the Claude 5.5 family, and he suspects the wording is partly aimed at investors. That is his opinion. A fair reading is that the letter sets expectations rather than announcing any change to Anthropic’s release plans.

Other IPO details mentioned

  • The prospectus lists the Nasdaq symbol “ANT” and an expected price range of $70 to $78 a share.
  • The video says pricing is expected on 18 November with trading from 19 November.
  • WorldofAI quotes $17 billion in quarterly revenue, more than 12 times higher year on year, and a large net loss mostly from accounting adjustments. Check these against the prospectus before relying on them.
  • The founders have pledged 80% of their personal shares to charity, and 5% of equity is set aside for a public-benefit fund.
  • Anthropic says Claude helped get one of its models running on AMD chips in 3 days with little human help, work it says used to take engineers months. That is the company’s own claim.

⚠️ Warning: The creator says he plans to invest in the Anthropic IPO. That is his personal view and he says it is not financial advice. Nothing in this article is investment advice either.

For background on the filing itself, see our earlier post on GPT-6.1 Terra and Anthropic’s IPO filing.

What is Claude Meetings?

Claude Meetings lets Claude join your video calls as a visible participant, answer questions, take notes and build documents or whiteboards while you talk. Anthropic launched it in beta on 10 November for Team and Enterprise plans, starting with Google Meet.

  • Claude appears in the participant list and posts a notice in the call’s chat when it joins.
  • After the call, the notes, action items, transcript and anything Claude made are saved on a meeting page.
  • Anthropic says it saves the transcript and screen descriptions, not audio.
  • During the beta, notes and transcripts are free on Team and Enterprise; documents and whiteboards made in calls are free until 10 February 2027.
  • Press reports say it will not join calls with external guests unless an admin allows it.

WorldofAI calls it a game-changer for team productivity. It may well be, but an AI that listens to every meeting needs clear rules about consent, what it can access and who checks its notes. Our practical guide How to Set Ground Rules for an AI Meeting Assistant gives you a one-page policy you can adapt.

📘 Main article: Before you switch Claude Meetings on, read How to Set Ground Rules for an AI Meeting Assistant. It covers consent, connected tools, approvals and a ready-to-copy team announcement.

Two smaller Anthropic items from the video are unconfirmed by us: invites for a Claude programme the creator calls “Duet”, said to include a free month of Claude Pro for eligible US users, and a private beta of self-hosted Claude Managed Agents for enterprises that want to run agents inside their own infrastructure.

What happened at NVIDIA?

NVIDIA has confirmed that Jensen Huang resigned as president, CEO and board member, effective immediately, and that CFO Colette Kress is interim CEO. NVIDIA’s announcement and its regulatory filing say the board accepted his resignation on 11 November 2026. Board member Harvey C. Jones became NVIDIA’s first-ever board chair, and the board is searching for a permanent CEO.

The video’s title says NVIDIA “fires” Jensen. NVIDIA describes it as a resignation, and its statement does not give a reason. Major news outlets have linked it to reports about his personal relationships, including national-security concerns raised in those reports. Those are allegations reported by the press, and we are not going to repeat the personal details here. Note that the video’s description of the allegations differs from mainstream reporting, so rely on established news outlets rather than social media summaries.

The Blackwell GPU report

WorldofAI spends more time on a separate story: a forensic analysis claiming that six recent Chinese models, from labs including DeepSeek, Moonshot AI, MiniMax and Ant Group, carry numerical “fingerprints” suggesting NVIDIA Blackwell chips were used when the models were first set up, despite US export restrictions. The researchers say they found the patterns in the publicly released model weights.

🔎 Unconfirmed / rumour: This is a research claim, not established fact. As the creator stresses, even if the hardware match is right, it does not show who supplied the chips, where they were used or whether any law was broken. There is no established link between this report and Jensen Huang’s departure.

Why it matters to you: most AI tools you use run on NVIDIA hardware, so leadership uncertainty and export pressure could affect prices and availability over time. Our article on the NVIDIA China chips standoff explains the knock-on effects for everyday users. NVIDIA reports its next quarterly results on 17 November.

What is OpenAI’s new cyber model?

During its “Cyber at the Frontier” presentation, OpenAI showed an unreleased internal cybersecurity model scoring 77% on ExploitGym, a test of whether an AI agent can turn a known software weakness into a working exploit.

ModelExploitGym score (as shown)
OpenAI’s unreleased cyber model77.0%
Gemini 4 Carbon54.8%
GPT-6 Cyber52.6%
GPT-6 Astra42.4%
Claude Mythos 5.2about 39%

These are the numbers WorldofAI read from the presentation. Note that OpenAI’s own GPT-6 Cyber launch page, from 5 November, lists a lower ExploitGym score of 48.5%, so the testing setups differ. The model has no name, release date or system card. The creator hopes it will reach verified security researchers through OpenAI’s Daybreak programme.

⚠️ Warning: The same skills that help defenders find weaknesses could help attackers. That is why these models are released slowly, if at all. For a small business, the practical step is unchanged: keep software updated and turn on two-factor sign-in everywhere.

ChatGPT Dots gets custom instructions

OpenAI also updated Dots, its personal AI agents in ChatGPT, with faster, plainer replies and better follow-through. The useful part is custom instructions, including rules for when a Dot must ask your approval, such as “always show me an email draft before sending”. WorldofAI says OpenAI also reset usage limits for everyone after people expected GPT-6.2 releases that did not arrive.

What else happened in AI this week?

StoryWhat was saidStatus
Claude Fable 5.2 animated shortA creator directed Fable 5.2 to make a 4-minute animated film by writing Python that drove Blender, with 6,000 frames rendered overnight on a single RTX 3070Creator’s showcase
Seedance 2.5 FlashByteDance’s faster, cheaper video model: about twice as fast as the flagship, around 15 cents a second at 720p (roughly 34% cheaper); API due 15 NovemberAnnounced
Handoff² by HarkA computer-use model said to be twice as fast as its predecessor and faster than Opus 5.5 at high reasoning, with better CAPTCHA and browser-task resultsCompany claims
Microsoft IgniteStarts 17 November; Microsoft’s MAI team says it will show new models and demosUpcoming
Gemini on Apple CarPlayiPhone users can talk to Gemini Live in the carRolling out

The Seedance price in the video was misspoken, so work it out yourself: at around 15 cents a second, a 10-second clip costs about $1.50.

What is confirmed and what is only claimed?

ClaimStatusHow we checked
Gemini 4 Hydrogen launched with 2.5x speed and $1.50 / $7.50 pricingConfirmedGoogle’s announcement and model card
Hydrogen ranks 7th on WoAI BenchCreator’s own testNot independently checked
Anthropic may advance more slowly for safetyConfirmed wordingAnthropic founders’ letter
Anthropic is slowing down nowNot claimed by AnthropicHeadline framing only
Claude Meetings beta on Google MeetConfirmedAnthropic’s announcement
Jensen Huang resigned; Colette Kress interim CEOConfirmedNVIDIA statement and filing
Reasons for his departureAllegations in press reportsNVIDIA gives no reason
Blackwell chips used by Chinese labsResearch claimNot proven; no link to the CEO change
OpenAI cyber model scores 77%Shown in a presentationNo system card or release
Claude “Duet” invites and free Pro monthUnverifiedCreator’s report only

💡 Tip: A good habit for any AI news: separate what a company has published from what someone says about it. Our checklist for spotting a fake AI model leak works for rumours in general, not just leaked models.

What should you do about this week’s news?

  1. Test Gemini 4 Hydrogen on one real task at the default setting before you change tools or plans.
  2. Check accuracy yourself if you rely on AI for facts, figures or documents. Use our made-up answers test.
  3. Agree meeting rules first if your team uses Claude Team or Enterprise and wants to try Claude Meetings. Start with our ground rules guide.
  4. Ignore the drama for decision-making. NVIDIA’s leadership change will not change the AI tools you use this week.
  5. Don’t base investment decisions on YouTube videos, including this round-up. Read the official documents and get proper advice.

Frequently asked questions

Is Gemini 4 Hydrogen free?

You can try it free in Google Search and, according to Google, in the Gemini app when automatic model selection chooses it. AI Studio and the API have free and paid tiers; paid API use costs $1.50 per million input tokens and $7.50 per million output tokens.

Is Gemini 4 Hydrogen the same as Gemini Flash?

It replaces Gemini 3.8 Flash as Google’s fast, lower-cost model. Google has moved to element names for Gemini 4, with Hydrogen in the role Flash used to fill.

Which Gemini 4 Hydrogen thinking level should I use?

Start with the default, high. WorldofAI found it gives the best balance of quality and token use; xhigh nearly doubles output tokens for a small improvement.

Is Anthropic slowing down AI development?

Not as an announced plan. Its founders’ letter says it is willing to advance more slowly than it could when safety requires it.

Who can use Claude Meetings?

Claude Team and Enterprise plans, in beta, starting with Google Meet. On Team plans it is on by default; Enterprise admins must switch it on.

Was Jensen Huang fired?

NVIDIA says he resigned and the board accepted his resignation on 11 November 2026. The company has not given a reason.

Was the video sponsored?

Yes, by Abacus AI (Abacus AI SuperComputer). The creator says he is not sponsored by Hark, whose Handoff² model he covers.

The bottom line

Gemini 4 Hydrogen is the most useful launch of the week for most people: fast, strong with documents and images, and priced the same as the Flash model it replaces. It is not the cheapest option and not the best coder, so test it on your own work before switching. The rest of the week’s headlines matter, but mostly as context. Check what has been confirmed, treat the rest as claims, and keep your decisions tied to your own results.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

Leave a Reply