Claude Mythos 5.5 Leak: Fable 5.5 Demos, GPT-6.2 Astra and This Week’s AI News

Every week brings a fresh batch of AI model leaks, and this one is busier than most. Anthropic, OpenAI and Google all appear to be lining up their next big releases, while the rules on how much you can use these tools are quietly changing.

Here is our plain-English summary of the Claude Mythos 5.5 leak and the rest of WorldofAI’s round-up, with what is confirmed, what is rumour, and what it means for how you work.

What is Claude Mythos 5.5, and is Fable 5.5 close?

Claude Mythos 5.5 appears to be the next version of Anthropic’s cybersecurity and biology research model class. According to the video, the name turns up six times in the HTML of Anthropic’s website, in hidden headings and image descriptions, even though the visible page still refers to Mythos 5.1.

An archived snapshot from 23 October reportedly didn’t contain those references, which suggests the site was updated in the days since.

Fable 5.5 has its own trail. WorldofAI says a model-tracking project briefly listed it as generally available in Azure’s East US 2 catalogue on 13 October, then removed it a day later when deployment requests returned errors. Another model list still shows it with a 1 million token context window at $10 per million input tokens and $50 per million output tokens.

The creator also reads Anthropic’s recent plan changes and its “more things cooking” message as hints that a launch is near. That is a fair guess, but still a guess.

What did the Fable 5.5 coding demos show?

The demos suggest Fable 5.5 can build large, working projects from a single prompt. WorldofAI shows a Fable 5.5 checkpoint, running in Claude Code at extra-high reasoning effort, producing a Minecraft-style game in one single file.

It reportedly took around one to two hours and included survival and creative modes, crafting, furnaces and even a second dimension. Not everything worked; the creator couldn’t get a portal to function.

These are selected clips from testers, not benchmark results. They show what is possible on a good run, not what you will get every time.

What is GPT-6.2 Astra, and what did OpenAI tease?

GPT-6.2 Astra is a model name some users report seeing in their Codex usage analytics, even though the model picker still shows GPT-6 Astra. WorldofAI reads this as OpenAI quietly testing the newer model with some users.

The early demos he shows are mixed. A PS5 controller drawn in SVG was decent but well behind the Fable 5.5 examples, while a 3D open-world RPG built in one go was more impressive.

OpenAI also posted a teaser video that it says was coded by AI. Viewers spotted scrambled text for about half a second that reportedly reads “Optimism Achieved Internally”, spelling out OAI. A chart shared alongside reportedly showed an unreleased internal model scoring about 26% on hard maths problems at its lowest compute, against 17% for GPT-6 Astra at its highest, and nearly 47% at higher compute.

Two smaller OpenAI items matter more if you work in those areas. A GPT-6 Astra for law listing reportedly appeared on the API pricing page at $10 per million input tokens, $50 per million output and $1 for cached input. And a notice says that from 29 November, conversations with its Daybreak Blue security model on personal accounts may be used for training even with the training setting switched off. Business and enterprise accounts are said to be unaffected.

What are Gemini 4 Xenon and Carbon-H?

Gemini 4 Xenon is a rumoured next flagship model from Google, said to be close to the ceiling on several benchmarks and cheaper to run. WorldofAI is open about not knowing how credible the reports are, and some of the claims involve Google using AI to speed up its own model development.

Carbon-H is easier to see. It is reportedly an improved Gemini 4 Carbon being tested in Arena under the name Argon 4. Demos include a Minecraft Dungeons-style game, a 3D Greek world and an animated superhero scene. In one head-to-head, Carbon drew a better PS5 controller than Opus 5.5, although the creator notes Opus was run at a lower reasoning level.

The point for most people is price. Carbon is expected to cost roughly half as much as Opus 5.5, so if the quality holds up it could be the better-value choice for coding and web work. We covered Carbon’s earlier checkpoint in our Gemini 4 Carbon round-up, and you can try unreleased models yourself with Arena battle mode.

What is changing with ChatGPT and Claude usage limits?

From December, ChatGPT Plus, Pro and Business users will reportedly share one usage allowance across ChatGPT and Codex, measured by tokens. Bigger models and longer reasoning will use up your allowance faster, wherever you use them. Instant chat is said to be excluded.

If you run coding agents in Codex all day, this could be a noticeable change. When you run out, the options are to wait for the reset, buy credits or upgrade.

Anthropic has made its own changes. The Claude plans have new names, and the headline benefit sits with the top tier:

There is a catch. The weekly limit on Ultra is 10 times the Pro allowance, not 20 times as many people assumed. Anthropic reportedly apologised for the confusing old names, and new accounts keep temporary limits until they are verified.

Why is Claude Code refusing to recreate existing games?

Claude Code reportedly received a server-side instruction telling it to create original games and not recreate a specific existing one, even if you rename it. WorldofAI says it arrived under an internal feature flag, and there is no setting to turn it off.

It doesn’t appear to ban game development, modding or games inspired by a genre. It targets direct copies. Some developers are reportedly moving game work to Codex instead, which is an understandable reaction given how many clone demos have done the rounds.

Our view: this is mostly Anthropic reducing copyright risk, and building something original was always the safer route if you plan to publish or sell. We wrote a step-by-step guide on how to build an original game with an AI coding agent. Last week’s reverse engineering block round-up has more background.

Was the Claude Fable 5 weights leak real?

No. The supposed Claude Fable 5 weights were essentially random numbers, not a trained model. An account on a model-sharing site claimed to have pulled the weights from a misconfigured cloud storage bucket, showed a 7 GB “shard” of a model said to total around 8 TB, and offered the files for sale.

The parameter count claims didn’t add up, and the files didn’t behave like real weights. As WorldofAI points out, fake leaks like this are a classic way to get people to download harmful files or run malicious code.

What else happened in AI this week?

• Project Suncatcher: Google reportedly sent a prompt to a TPU on a test satellite and got a haiku back, an early proof that AI hardware can run in orbit.
• Anthropic safety research: in a deliberately fine-tuned test model, researchers found worrying behaviour, and used thousands of Claude agents to create examples that slip past safety monitors so the monitors can be improved.
• OpenCode: you can reportedly connect your Claude subscription to OpenCode through a new Claude Agent integration, without paying separately for API use.
• Google AntiGravity: agentic coding is now inside Android Studio, a “Chief of Staff” agent is reportedly in development, and Jules is reportedly being folded in.
• Video and images: Google’s Omni Flash 1.2 video model may support 1080p and 4K, and a new Seedance 5.0 family is reportedly close.
• Other models: GLM 5.5 is said to be in advanced post-training, and MiniMax M3.1 Pro is being tested in Arena.
• OpenAI’s 28-day shipping run ended with account switching and a usage reset rather than a new model.
• Local AI debate: Sam Altman’s comments about personal computer memory sparked backlash from people who want to run AI on their own hardware.

The creator also suggests keeping a capable local model and a backup power source for emergencies. That is a personal choice, but having an AI tool that works offline is useful for anyone who travels or has patchy internet.

What should you actually do this week?

1. Don’t change your stack on a leak. Note Fable 5.5, GPT-6.2 Astra and Gemini 4 Carbon as “test when released”.
2. Check your plan. If you are on ChatGPT Plus, Pro or Business and use Codex heavily, look at your usage before December.
3. Review Claude Ultra’s real limit. It is 10 times Pro per week, so plan long agent jobs around that.
4. Never download “leaked” closed models. Use our fake leak checklist if one appears in your feed.
5. Keep game and app projects original. Follow our original game guide if you use AI coding agents for games.

Frequently asked questions

The bottom line

The Claude Mythos 5.5 leak and the Fable 5.5 demos point to a big Anthropic release soon, with OpenAI and Google clearly getting ready to answer. But the changes that affect your work today are the less exciting ones: shared ChatGPT and Codex limits, Claude’s new plan names, and a firmer line on copying existing games. Your next step: check your current AI usage this week, so a new limit or a new model doesn’t catch you out.

Related posts

• Gemini 4 Carbon: Google’s New Frontier Model, the Fable 5.5 Leak and This Week’s AI News
• GPT-6.2 Bel Leak: OpenAI’s Next Model, Opus 6 Testing and This Week’s AI News
• What to Do When Your AI Plan’s Usage Limits Are Cut

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.

See Full Bio