Quick answer
The latest Claude Fable 5.5 leaks point to a release that could be close, but Anthropic has not announced anything. Two new early-access model names, claude-orzata-eap and claude-vanilla-eap, were reportedly found inside Devin’s CLI, and leakers say Anthropic’s partners already have access to Fable 5.5. The “baby AGI” label comes from impressive single-prompt demos, such as a Minecraft-style game in one HTML file. They are striking, but they are not proof of general intelligence. The confirmed news this week is more useful: Claude responds faster, Claude Mythos 5 designed better quantum circuits in Anthropic’s own research, and GPT-6 Astra Pro can now run parallel agents.
Key takeaways
- Leaks (unconfirmed): new Claude EAP names in Devin’s CLI, Fable 5.5 reportedly listed on Azure, and partner access. A release next week is speculation
- Confirmed by Anthropic: Claude Code now compresses each turn with gzip, and cloud sessions start as you type. Reported start-up times are up to 3x faster
- Confirmed research: Claude Mythos 5 worked in a closed loop with an automatic checker and beat the best published designs on eight quantum-circuit tasks, including 17x fewer gates on one
- OpenAI: GPT-6 Astra Pro gets faster responses, better tool use and parallel agents for Pro users. Codex and ChatGPT Work stretch usage limits further
- Rumours: GLM 5.5 (reportedly over 1 trillion parameters) and Kimi K3.1 may land in October. Treat the benchmark claims with caution
In this guide
- Watch: the source video
- What are the latest Claude Fable 5.5 leaks?
- Is Claude Fable 5.5 really “baby AGI”?
- What did Claude Mythos 5 do with quantum circuits?
- How much faster is Claude Code now?
- Is GLM 5.5 about to beat Fable and GPT-6?
- What changed in GPT-6 Astra Pro and Codex?
- What else is new: Gemini 4 Argon, free models and a giant robot
- What should small businesses do with this news?
- Frequently asked questions
- The bottom line
There are more Claude Fable 5.5 leaks this week, and the hype has a new label: “baby AGI”. Underneath the excitement there are a few real, useful changes you can use today.
Below, we separate what has been confirmed from what is only rumoured, and explain what each story means if you run a small business or build automations. For the earlier rounds of leaks, see Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update and Muse Spark Open: Meta’s Open-Weight Model, Opus 6 Rumours and This Week’s AI News.
🛠️ Practical guides
Want to act on this week’s news? These step-by-step guides answer two follow-up questions:
- How to Split a Task Between Parallel AI Agents Without Making a Mess — how to divide one job into independent pieces, brief each agent and merge the results safely.
- How to Build an AI Agent Feedback Loop: Brief, Check, Score, Repeat — how to copy the write, check and improve loop from Anthropic’s quantum research for everyday business tasks.
Watch: the source video
What this video covers: World of AI (@intheworldofai) rounds up new Claude Fable 5.5 leaks and early-access checkpoints, eye-catching demos from “Claude Next”, GLM 5.5 and Kimi K3.1 rumours, Claude’s speed upgrade, Claude Mythos 5’s quantum-circuit research, OpenAI’s Codex and GPT-6 Astra Pro updates, Gemini 4 Argon in Google AI Studio, free stealth models in OpenCode and a 13-foot humanoid robot from China.
Video: World of AI (@intheworldofai). This summary and the opinions in it are our own. Leaks and rumours are labelled as unconfirmed.
What are the latest Claude Fable 5.5 leaks?
Short answer: several small signals all point the same way, but none of them comes from Anthropic. Here is what has been reported since our last round-up:
| Signal | What was reported | How much weight to give it |
|---|---|---|
| Azure model list | Fable 5.5 reportedly appeared in Microsoft Azure’s model list earlier in the week. The release many expected that week did not happen | A useful hint, but it already proved a poor guide to timing |
| Devin CLI binary | Two new code names, claude-orzata-eap and claude-vanilla-eap, were spotted in the Devin coding agent’s software | Suggests checkpoints are being tested. It doesn’t say which models they are |
| Earlier EAP names | “Verrine” and “Onigiri” were reportedly added about a week ago. Onigiri is the name linked to Fable 5.5 in last week’s leaks | Fits the pattern, still unconfirmed |
| Partner access | Leakers described as reliable say Anthropic partners recently got access to Fable 5.5, so a release next week looks likely to them | Second-hand. Treat it as a maybe |
| “Claude Next” routing | Several checkpoints appear to sit behind a “Claude Next” option that some people can reach in Claude Code | Hard to verify from the outside |
🔎 Unconfirmed / rumour: Four code names does not necessarily mean four models. Labs often test several checkpoints of the same model. Until Anthropic publishes a model card, pricing and availability, Fable 5.5 is a rumour, however many names leak.
If you want a quick way to judge leaks like these, our guide on how to verify AI model leak rumours gives you a ten-minute check.
Is Claude Fable 5.5 really “baby AGI”?
Short answer: no, and the creator says so too. He calls it “baby AGI” as shorthand for models that can turn a short prompt into a detailed, interactive world, rather than as a literal claim. The demos behind the label are worth knowing about:
- VoxelBench comparison. On one voxel-art prompt, Fable 5.5 reportedly built a connected 3D scene with lighting that you can explore. Opus 5.5’s version looked clean but flat and static.
- A Minecraft-style game in one HTML file. Using “Claude Next” on medium effort, a tester asked for a Japanese-style sandbox game. The result had floating blocks, sand that falls, torches that go out underwater, trees that fall, background music, wind sounds and a dolphin that starts doing tricks.
- A Tesla Roadster launch video rebuilt in code. Another tester gave the model a reference clip and had it recreate the scene using Python inside Blender, with ElevenLabs for audio.
🔎 Unconfirmed / rumour: All three demos were shared by testers on X. We can’t tell which model or checkpoint produced them, how many attempts were made, or what they cost. Single-prompt showcases are usually the best results, not the average ones.
What the demos do show is progress in spatial reasoning and in filling gaps sensibly: adding physics, sound and small details the prompt never asked for. That is exciting for anyone building games, product mock-ups or explainer animations. It still isn’t the same as a model that understands the world or works reliably on your tasks.
💡 Tip: When Fable 5.5 does launch, judge it on three or four of your own real jobs, not on demo videos. Our guide to building an AI test set shows how to prepare those tests now, so you are ready on launch day.
What did Claude Mythos 5 do with quantum circuits?
This one is confirmed. On 16 October Anthropic published research showing Claude designing more efficient quantum circuits. Claude Mythos 5 was given plain-English descriptions of eight circuits used in chemistry, optimisation, physics simulation and number theory, plus an automatic evaluator that checked each design and scored it.
The headline results from Anthropic’s write-up:
| What Anthropic reported | Detail |
|---|---|
| Better than published designs | Across all eight tasks, Claude found designs that used fewer qubits and fewer gates than the best previously published circuits |
| Elliptic-curve point addition | One new design uses 12% more qubits but 17x fewer gates than the smallest previous design |
| Quantum optimisation decoder | The most expensive step of a leading algorithm became about 12x cheaper than Google’s previous construction |
| How long it worked alone | One run lasted 53 hours from a one-paragraph brief, with only occasional prompts to think more broadly |
| Limitations | Claude tended to be cautious, double-checking good results and preferring well-known approaches. Asking it to try several different approaches in parallel helped |
The video quotes slightly different figures, including a 39x improvement on one circuit. We have stuck to the numbers in Anthropic’s own post.
The part that matters for most of us is how it worked. Claude proposed a design, an automatic checker tested and scored it, and Claude used that feedback to improve the next attempt, with no human in between. You can use the same pattern for far more ordinary jobs, such as meta descriptions, product listings or spreadsheet formulas. Our guide How to Build an AI Agent Feedback Loop: Brief, Check, Score, Repeat shows you how.
How much faster is Claude Code now?
Short answer: Claude now starts responding sooner, especially in Claude Code. Anthropic’s developer account confirmed two changes: Claude Code now compresses (gzips) the conversation it sends on every turn, and cloud sessions start up as soon as you begin typing.
| Product | Reported improvement |
|---|---|
| Claude Code, local sessions | About 3x faster to start responding (roughly 4 seconds down to 1.33 seconds) |
| Claude Code, cloud sessions | About 2.8x faster (roughly 17 seconds down to 6 seconds) |
| Claude chat and Cowork | About 1.5x faster |
| Claude API | About 1.5x faster time to first token |
⚠️ Warning: These gains cut the wait before an answer starts. They don’t make the model write three times faster. If you need faster output as well, Opus 5.5 Fast Mode (now generally available) is up to 2.5x faster, at roughly twice the price.
The video also highlights Claude’s ability to build slide decks inside artifacts, with layouts, visuals and device mock-ups. It isn’t new, but it is worth knowing about if you make sales or training decks.
Like practical takes on AI news? Each week we turn the headlines into clear next steps for small businesses. Browse the latest guides on Prodify Digital.
Is GLM 5.5 about to beat Fable and GPT-6?
Short answer: nobody knows yet. The claims are bold and come from unnamed sources.
🔎 Unconfirmed / rumour: Several sources claim GLM 5.5 from Z.ai could outperform Anthropic’s current Fable model and OpenAI’s GPT-6 Sol. It is rumoured to have over 1 trillion total parameters and a 1 million token context window. There is also speculation that a GLM 5.5 Flash variant has already appeared in OpenCode, and strings spotted in model catalogues hint at GLM 5.5 and Kimi K3.1 launching in October. None of this is confirmed.
If GLM 5.5 does arrive as an open-weight model at that level, it would put real pressure on closed-model prices. For now, the sensible move is to keep your current tools and add GLM 5.5 to your test list for launch day.
What changed in GPT-6 Astra Pro and Codex?
OpenAI is now on day 12 of its 28-day shipping pledge. Here is what is new:
- More from your Codex and ChatGPT Work limits. Context is now reused across sub-agents and cached for longer, and requests are routed to hit that cache more often. Less repeated processing means your usage goes further, especially on multi-agent jobs.
- GPT-6 Astra Pro upgrade. In ChatGPT, Astra Pro now responds faster, handles plugins and tools better and, most importantly, can run parallel agents. Several agents work on different parts of a task at once, and you can track each one and open its finished work. It is available to Pro users in the chat tab.
- Slides in ChatGPT Space. Space, launched at DevDay, now supports presentations. It is available on Pro, Business and Enterprise while OpenAI gathers feedback.
- Dot widget on iOS. You can add your Dot to your iPhone home screen and start a conversation or call with one tap.
🔎 Unconfirmed / rumour: One user reports that Astra Pro’s parallel agents don’t use your Codex quota, and that Pro mode allows about 50 messages a week on the $100 plan and 100 on the $200 plan. OpenAI hasn’t confirmed these figures. Whether a GPT-6.1 Astra Pro is coming soon is pure speculation.
Parallel agents are powerful, but only if the task splits cleanly. Our guide How to Split a Task Between Parallel AI Agents Without Making a Mess covers how to divide the work, brief each agent and merge the results without a mess.
What else is new: Gemini 4 Argon, free models and a giant robot
- Gemini 4 Argon in Google AI Studio. Google AI Pro subscribers can now build with Argon in AI Studio’s Build mode and Playground. If you already pay for Pro, it is an easy way to get more value from it. Our guide to Google AI Pro vs Ultra for Gemini 4 Argon explains the plans.
- Devin offers. According to one user, Devin has extended free access to its SWE 2 model until the end of the month for subscribers, and Gemini 4 Argon is 75% off in Devin’s CLI and desktop app. Check the offer on Devin’s own site before you sign up.
- A stealth model to skip. A free model called “LL Alpha” appeared in OpenCode. Some guessed it was an upcoming GLM model. In one test it spent about 40 minutes thinking and another hour working, produced an HTML file that showed nothing, then insisted its code was right and blamed the tester. The advice for now is to avoid it.
- GLM 5.3 Flash is free in OpenCode. It is a better free option than LL Alpha for experiments.
- A 13-foot humanoid robot. Beijing-based NeodraDynamics showed Neo1, a humanoid about 4 metres tall weighing around 1.2 tonnes. It is reportedly rated to lift about 500 kg (over 1,100 lb) with both arms, and is designed for dangerous industrial work such as heavy assembly, ports and mines. It is reportedly heading for trials in JD Industrial’s supply chain. Strength is not intelligence, and these figures are the company’s claims.
⚠️ Warning: Free stealth models are fine for curiosity, but never give them client data or work that matters. You often don’t know who runs them or what happens to your prompts. Set a time limit on any test, so one stubborn model can’t eat two hours of your day.
What should small businesses do with this news?
| Story | Status | What we would do |
|---|---|---|
| Claude Fable 5.5 | Rumour (leaks, partner access) | Prepare a test set now. Switch nothing until launch |
| Claude speed upgrade | Released | Nothing to do. Your Claude Code sessions should feel snappier |
| Mythos 5 quantum research | Published research | Borrow the feedback-loop pattern for your own repeatable tasks |
| GPT-6 Astra Pro parallel agents | Released (Pro) | Try it on one research task that splits into independent parts |
| Codex and ChatGPT Work limits | Released | Re-check how far your plan goes before you upgrade |
| GLM 5.5 and Kimi K3.1 | Rumour | Add to your watch list |
| Gemini 4 Argon in AI Studio | Released (AI Pro) | Use it if you already pay for AI Pro |
| LL Alpha stealth model | Available, poor early tests | Skip it for now |
The pattern is the same as every week. Act on what has shipped, test it on your own work, and keep rumours in a “watch” column. If you are juggling several models already, our guide to AI Model Routing for Small Business: Match Each Task to the Right Model helps you match each task to the right one.
Want to turn this week’s news into a working system? Start small: split one research task between parallel agents, or build a feedback loop that checks an agent’s work for you, using our step-by-step guides.
Frequently asked questions
Has Claude Fable 5.5 been released?
No. As of 17 October 2026 Anthropic has not announced Fable 5.5. Leakers say partners have access and expect a release next week, but that is unconfirmed.
What are claude-orzata-eap and claude-vanilla-eap?
They are early-access model names reportedly found in the Devin CLI software. They are probably test checkpoints of upcoming Claude models, but nobody outside Anthropic knows which models they are.
Is Claude Fable 5.5 AGI?
No. “Baby AGI” is the creator’s shorthand for impressive demos where one prompt produced a detailed interactive game or animation. The demos show progress in spatial reasoning and detail, not general intelligence.
What did Claude Mythos 5 achieve in quantum computing?
In Anthropic’s research, Claude Mythos 5 designed quantum circuits that beat the best published designs on eight tasks. One used 17x fewer gates and another step became about 12x cheaper. It worked in a loop with an automatic checker.
Why does Claude Code feel faster?
Claude Code now gzips the conversation it sends each turn, and cloud sessions start as you type. Reported start-up times are about 3x faster for local sessions and 2.8x for cloud sessions.
What are parallel agents in GPT-6 Astra Pro?
They let several AI agents work on different parts of one task at the same time inside ChatGPT, with progress you can track. The feature is available to Pro users.
When will GLM 5.5 come out?
Rumours point to October 2026, alongside Kimi K3.1, but Z.ai has not confirmed a date, its size or its benchmark results.
The bottom line
The Claude Fable 5.5 leaks are getting louder, and the demos are genuinely impressive, but there is still no launch, pricing or model card. The news you can use today is quieter: Claude starts responding sooner, GPT-6 Astra Pro can split work across parallel agents, and Anthropic’s quantum research shows how powerful a simple write, check and improve loop can be. Prepare your tests, keep rumours on your watch list and put the confirmed features to work.
Related posts
- Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update
- How to Verify AI Model Leak Rumours Before You Switch Tools
- Muse Spark Open: Meta’s Open-Weight Model, Opus 6 Rumours and This Week’s AI News
[…] This guide answers one question: how do you set up a loop where an AI agent checks and improves its own work? It was prompted by Anthropic’s quantum circuit research, covered in our round-up of the Claude Fable 5.5 leaks and this week’s AI news. […]