Low-poly glowing teal crystal on a plinth beside a second plinth hidden under a cloth, representing the Gemini Krypton leak and a hidden Fable model

Quick answer

Gemini Krypton is the reported code name of a new, unreleased Google Gemini checkpoint that early testers say beats Gemini 4 Carbon on visual and coding tasks. It was reportedly spotted as a model name in backend logs, and side-by-side demos suggest richer output, though it takes about twice as long on the same prompt. In the same week, Anthropic released Claude Fable 5.2 and Mythos 5.2, the long-rumoured Fable 5.5 did not appear, and OpenAI and Meta both had news too. Krypton is unconfirmed. Treat it as a sign of where Google is heading, not a reason to change tools today.

Key takeaways

  • Gemini Krypton is an early, unconfirmed checkpoint that reportedly outperforms Gemini 4 Carbon and Argon in demos, but runs slower.
  • Gemini 4 Hydrogen, believed to be part of the Flash family, is reported to be in private early testing.
  • Anthropic’s Claude Fable 5.2 is priced at $8 / $40 per million input/output tokens, about 20% cheaper than Fable 5.1, according to the video.
  • Fable 5.2 was reportedly held back on purpose on some AI self-improvement tasks, which has fuelled theories that a stronger model is waiting.
  • Meta’s Muse Ember is a small, cheap model at $0.10 / $0.50 per million tokens with a 1 million token context window.
  • OpenAI says its new Control Room system can stop a research model producing output within 35 seconds of a shutdown decision.
  • Code names, leaked checkpoints and IPO timing in this round-up are leaks or secondary reports unless stated otherwise.

🛠️ Practical guides

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

This was one of the busiest weeks of the year for AI news. Anthropic shipped a new flagship, Google appears to be testing yet another Gemini checkpoint only weeks after Carbon, and OpenAI and Meta both announced things that matter for anyone running AI in a small business.

This round-up focuses on the Gemini Krypton leak, then summarises the Claude Fable 5.2 launch, the missing Fable 5.5, GPT-6.2 Sol, Meta’s Muse Ember and OpenAI’s new kill switch. Our earlier coverage of Gemini 4 Carbon explains the model Krypton is being compared with.

Watch: the source video

What this video covers: WorldofAI’s 10 November AI news round-up covers Claude Fable 5.2 and Mythos 5.2 (benchmarks, pricing, demos and usage resets), why Fable 5.5 is missing, GPT-6.2 Sol in Codex, the Gemini Krypton and Gemini 4 Hydrogen leaks, Meta Muse Ember, free models in OpenCode, a Codex retro handheld, Anthropic’s IPO timing, OpenAI’s Control Room kill switch, Google Antigravity’s knowledge work interface and Tesla Optimus 2.5. The video is sponsored by Abacus AI.

Video: WorldofAI (@intheworldofai). Sponsored by Abacus AI (Abacus AI SuperComputer). This summary and the opinions in it are our own. Code names, leaked checkpoints and unreleased features are leaks or secondary reports until the companies confirm them.

What is Gemini Krypton?

Gemini Krypton is the code name of a new Gemini model checkpoint that has reportedly appeared as a model slug in backend logs. Google has not announced it. Everything we know comes from leakers and early side-by-side comparisons shared online.

  • It is described as a very early checkpoint, the first in a new series.
  • Testers say it is significantly better than Gemini 4 Carbon, which Google released only recently.
  • WorldofAI suggests Google now seems to be aiming for a major model update roughly every month. That is the presenter’s reading, not a Google statement.

🔎 Unconfirmed / rumour: Gemini Krypton is a leak. Its name, release date, pricing and final performance are all unconfirmed, and early checkpoints often change before launch. Before you act on any leak, run it through our checklist for spotting a fake AI model leak.

How does Gemini Krypton compare with Gemini 4 Carbon?

In the demos shown, Krypton produces more detailed visual output than Gemini 4 Carbon, but it takes about twice as long to answer the same prompt.

Test shown in the videoWhat reportedly happened
Side-by-side scene generationsKrypton’s output looked richer and more detailed than Carbon’s on the same prompts.
PS5-style controller as SVG codeKrypton’s controller looked cleaner and more accurate than Carbon’s.
“Pelican riding a bike” SVG testKrypton beat both Carbon and Gemini 4 Argon, but took noticeably longer.
Short filmA short film generated with Krypton was shown as an example of its first checkpoint.

The trade-off matters. A model that thinks for longer can give better results on creative and visual tasks, but it costs more time and usually more tokens. Our guide on choosing an AI reasoning effort level explains how to decide when slower is worth it.

Gemini 4 Hydrogen is also back in the news. A leaker WorldofAI rates highly says it has entered private early testing with a limited group, which often comes before a wider release. Hydrogen is believed to belong to Google’s faster, cheaper Flash family. Again, this is unconfirmed.

What did Anthropic announce with Claude Fable 5.2?

According to WorldofAI, Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, two models aimed at hard, long-running technical work, and both reportedly beat Claude Opus 5.5 on nearly every capability Anthropic tested. For a deeper look at how well it performs, read our full Claude Fable 5.2 review.

The headline points from the video:

ItemWhat was reported
SpeedOutputs up to 30% faster than the previous model, according to early testers.
Price$8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1.
EfficiencyOn Frontier Code 2.0, Fable 5.2 at high effort reportedly matched Opus 5.5 at max effort at roughly half the cost per task.
RankingsNumber one on the Artificial Analysis Intelligence Index with a score of 65, and narrowly first on WorldofAI’s own benchmark.
Fable vs MythosThe same underlying model. Mythos 5.2 has fewer restrictions and is only for verified organisations doing cyber security or biomedical work.
Daily useAnthropic still recommends Opus 5.5 as the everyday model for coding and knowledge work.
SubscribersBetween one and three free Fable usage resets depending on how long you have subscribed, and a weekly Fable allowance for Pro users equal to 50% of the normal weekly limit.

The demos are striking: an operating system built by more than 1,000 sub-agents over 8 days, three electronic devices designed from a one-paragraph prompt, and a game that consumed about 2.7 million output tokens and cost roughly $550 in API fees. In one side-by-side test, Fable 5.2 finished a 3D scene about 75 minutes faster than Opus 5.5 but cost about $81 against $46.

⚠️ Warning: Those costs are the real story for most businesses. Fable 5.2 is described as very token-heavy, especially at the highest effort settings. Set a budget before you try long jobs, using how to set an AI API spending limit and usage alert.

Anthropic also says the model is better at controlling physical robots, including on tasks it was not trained for. That is an unusual claim for a launch post and worth watching, but it does not change how most of us will use it.

Is Anthropic hiding a stronger Fable model?

Possibly, but nobody outside Anthropic knows. The rumoured Fable 5.5, which we covered in our earlier Fable 5.5 round-up, did not arrive. Instead, Anthropic shipped 5.2.

What the video points to:

  • Anthropic’s reported training details say Fable 5.2 was deliberately kept around Opus 5.5’s level on two AI self-improvement tasks: optimising AI chip code (kernels) and improving language model designs.
  • Other skills, such as agentic coding, hardware engineering and general machine learning research, were reportedly left unrestricted.
  • Fable 5.2 is believed to be an older checkpoint from a series of models that recently appeared in testing, such as those in our Claude Silicon leak article.

🔎 Unconfirmed / rumour: The idea that Anthropic is holding back its strongest model, perhaps a Fable 5.5, Fable 6, Opus 6 or Sonnet 6, until OpenAI makes its next move is the presenter’s theory. Limiting self-improvement abilities can also simply be a safety choice. Neither explanation is confirmed.

What is the latest on GPT-6.2 Sol?

GPT-6.2 Sol is an unreleased OpenAI model that some users, including WorldofAI, say they are being routed to inside Codex when they pick GPT-6.1 Sol with high thinking. The demo shown is a detailed 3D lake scene built with the Three.js library, with a boat, waterfall, trees and butterflies. We cover the early tests in more detail in our GPT-6.2 Sol preview article.

OpenAI has not announced GPT-6.2. If you think a tool is quietly serving you a different model, our guide on checking which AI model a provider is really serving you can help.

What is Meta Muse Ember?

Muse Ember is Meta’s smallest and most efficient model so far, built for reasoning and everyday AI agent work at a very low price. Unlike most items here, this is an official announcement.

  • A 1 million token context window, with support for text, images, audio and video.
  • Meta says it is close to Muse Spark 1.3 Max in intelligence at about 17 times lower cost.
  • Meta also claims it performs like Claude Haiku 5.5 at about a third of the cost per task, and beats GPT-6 Luna across reasoning settings.
  • It was trained using Meta’s larger Muse Wonder model. The idea is that Wonder plans and Ember handles many smaller tasks.
  • Priced at $0.10 per million input tokens and $0.50 per million output tokens, through the Meta Wonder API and Muse Code.

Small, cheap models like this are ideal for routine jobs such as tagging, sorting and first drafts. See how to route simple tasks to cheaper AI models for a practical way to use them alongside a frontier model.

💡 Tip: The performance comparisons are Meta’s own claims. Test Ember on three of your real tasks before moving any work across.

What is OpenAI’s Control Room kill switch?

Control Room is an emergency shutdown system OpenAI has introduced for its frontier AI research workloads. It monitors AI jobs across OpenAI’s research systems and can shut them down quickly if safety or security concerns come up.

  • It can stop a model producing further outputs within 35 seconds of a shutdown decision, even if the underlying experiment keeps running.
  • AI monitoring systems can trigger interventions when they spot worrying behaviour.
  • It is fail-closed: if Control Room crashes or is tampered with, all requests are blocked by default.
  • OpenAI says this is already active across its frontier research work, with wider coverage planned.

WorldofAI stresses that this does not mean OpenAI has found a rogue AI. It is a precaution. The lesson for small businesses is simple: if you run agents or automations, you should be able to stop them all quickly, too. Our new guide shows how to add a kill switch to your AI agents and automations.

What else happened in AI news this week?

  • Free models in OpenCode: MiniMax M3.1 Flash is currently free, and LongCat 2.5 is reportedly free for two weeks with a 1 million token context window and a zero data retention policy. Read how to use a free AI model without risking your data before you try them.
  • Google Antigravity for knowledge work: a new interface for research, writing and creating documents, using the same agent, plugins and MCP tools but without the terminal and Git clutter. Our guide on using an AI agent for research and writing shows how to get started with this kind of work in any agent tool.
  • Codex retro handheld: OpenAI has a special edition ModRetro Chromatic handheld and a Codex plugin to help you build, design and play-test your own retro games.
  • Anthropic IPO: Bloomberg reports Anthropic could start marketing its IPO as early as Tuesday, with a Nasdaq debut possibly next week. Reported valuations range from $1.1 trillion to $1.3 trillion. Our Anthropic IPO article explains why it matters.
  • Trump comment: President Trump reportedly said he had warned the unnamed leader of a major AI company that they could face arrest, and that the person then changed course. No arrest took place and the person was not named.
  • Tesla Optimus 2.5: footage shows the humanoid robot handing out snack packets to attendees at an event in Austin.
  • Sponsor: the video is sponsored by Abacus AI, promoting its SuperComputer tool for building 3D spaces, games and apps from prompts.

What does the video not confirm?

🔎 Unconfirmed / rumour: The video does not confirm that Gemini Krypton or Gemini 4 Hydrogen will be released, what they will be called or when. The Krypton comparisons are early demos shared online. GPT-6.2 Sol is unannounced and the routing reports come from users. The idea that Anthropic is holding back Fable 6 or a similar model is speculation. The IPO timing and valuation come from press reports. Benchmark rankings and Muse Ember’s comparisons come from the companies or third parties, and we have not tested them ourselves.

What should you do about this week’s news?

  1. Keep your current tools. Nothing here is a reason to switch today, especially the leaks.
  2. Budget before you try Fable 5.2 on long jobs. It is powerful but token-heavy. Compare models on token cost, not just output quality.
  3. Use cheap models for routine work. Muse Ember and free OpenCode models are worth testing on low-risk tasks with dummy data.
  4. Add an emergency stop to your automations, following our kill switch guide.
  5. Try an agent for non-coding work. Use our research and writing guide on one real task this week.
  6. Wait for official announcements before you plan around Krypton, Hydrogen or GPT-6.2.

Who should pay attention?

This week matters most if you build with AI coding agents, run automations that use AI, plan AI budgets for next year, or follow Anthropic’s IPO.

It matters less if you mainly use AI for everyday writing and email. The practical takeaways for you are the cheaper small models and the new focus on agents for research and writing.

Frequently asked questions

What is Gemini Krypton?

Gemini Krypton is the reported code name of an early, unreleased Gemini checkpoint spotted in backend logs. Early demos suggest it beats Gemini 4 Carbon on visual tasks, but Google has not confirmed it.

Is Gemini Krypton better than Gemini 4 Carbon?

In the side-by-side demos shown, yes, especially for detailed visual and SVG output. It reportedly takes about twice as long on the same prompt, and these are early, unverified tests.

How much does Claude Fable 5.2 cost?

According to the video, $8 per million input tokens and $40 per million output tokens, about 20% cheaper than Fable 5.1. It uses a lot of tokens at high effort, so real costs can still be high.

Where is Claude Fable 5.5?

It has not been released. Anthropic shipped Fable 5.2 instead, and reportedly limited it on some AI self-improvement tasks. Claims that a stronger Fable 5.5 or Fable 6 is being held back are speculation.

What is Meta Muse Ember?

Muse Ember is Meta’s small, low-cost model for reasoning and agent work, with a 1 million token context window, priced at $0.10 input and $0.50 output per million tokens.

What is OpenAI’s Control Room?

It is an emergency kill switch for OpenAI’s frontier research workloads. OpenAI says it can stop a model’s outputs within 35 seconds of a shutdown decision and blocks all requests if it fails.

Who sponsored the WorldofAI video?

The video is sponsored by Abacus AI, which makes the Abacus AI SuperComputer tool for building games and apps from prompts.

The bottom line

Gemini Krypton suggests Google already has a stronger model than Gemini 4 Carbon in testing, while Anthropic’s Claude Fable 5.2 raises the bar for long, complex work at a lower price per token. The missing Fable 5.5, GPT-6.2 Sol and Gemini 4 Hydrogen are still unconfirmed. For your business, the useful moves are practical: control your AI costs, try cheaper models on routine jobs, give your automations a clear emergency stop, and test agents on research and writing as well as code.

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