Published: October 3, 2026 | By Vito Ruocco
Anthropic’s next flagship might already be answering your questions. People who choose Fable 5.1 on Claude’s website have started getting answers that Fable 5.1 should not be able to give, and the word in the community is that Fable 5.5 could launch as early as next Tuesday. I wanted to know how much of that is real, so I did two things: I ran the community’s routing test myself through the API, and I gave the Fable you can actually buy today, and a model that costs absolutely nothing, the same four prompts people have been using to show off Fable 5.5.
This article is the long, written version of my latest VitoMind video. It covers the whole week in AI that led up to the Fable 5.5 leak: Claude Code becoming a moddable tool, Microsoft rebuilding Copilot, DeepSeek’s new desktop harness, a free 560-billion-parameter model on OpenCode, Gemini 4 Argon’s mixed first tests, a hidden Nano Banana string in a Google app, and the true story behind the most impressive Fable 5.5 demo of all, a short film about the Hugging Face incident. Then we get to the part that matters most: what happened when I put Fable 5.1 and a free model head to head, with every demo published so you can use it yourself.
If you prefer to read, everything from the video is below, with more detail than I could fit into eleven minutes.
The Short Version
Before we go deep, here is the summary for anyone in a hurry.
Fable 5.5 is not officially announced. Anthropic has not published a model card, a price or an API identifier for it. The last two official releases were Claude Opus 5.5 on September 22 and Claude Sonnet 5.5 on September 28.
But something is being served on claude.ai. Users who pick Fable 5.1 on the website are getting answers that show knowledge Fable 5.1 does not have. A simple one-line test, which I explain below, makes the difference visible, and when I ran it through the API, the API version of Fable 5.1 failed it.
The free model surprised me. On four prompts modelled on the Fable 5.5 showcase demos, Fable 5.1 cost about ten dollars and forty minutes. Ling 3.1 Flash cost nothing and took about eleven minutes. On the two games, the free model was honestly close. On the animated film and the design app, Fable was clearly ahead.
You can try every demo. All eight results are live on my site, free to use in your browser: ruocco.it/ai-demos/fable-5-1-vs-ling-3-1. The exact prompts and the full source code are in the Ruocco Academy.
Claude Code Now Takes Mods
Let’s start with something you can use today. Anthropic has turned Claude Code into a moddable tool.
The official announcement describes it in three lines: a mod can change how Claude Code behaves, customise its interface, or swap in your own features. You can write one in a few lines of TypeScript, or you can simply ask Claude to write it for you. Mods ship inside plugins, so you install them with the /plugin command in the CLI or the desktop app.
The official guide, published on October 1 and written by Addy Osmani, explains the idea clearly. A mod is a small JavaScript or TypeScript file that runs inside your Claude Code session. It can watch what is happening, change what Claude Code does, or draw its own interface, in the terminal or in the desktop app. Under the hood, mods are hooks that ship inside plugins, and each one sees every event in your session as it happens: tool calls, the prompt you submit, turns starting and finishing, slash commands, and every piece of the interface as it is drawn.
Observe, rewrite, or answer
What I like most about the design is its simplicity. Every hook works like middleware, and it can do one of three things.
It can observe: let the event pass and just look at the result, for example recording every file edit or taking a reading after each turn.
It can rewrite: change the event before the rest of the chain sees it, for example replacing a risky command with a safer one.
Or it can answer: stop the event right there and respond itself, for example refusing a tool call or serving a command on its own.
Token Weather and Blast Radius
The first example in the guide is my favourite, because it solves a problem every heavy Claude Code user has. It is called Token Weather, and it draws a forecast of your context window in a band above the prompt. Under 25 percent full, the forecast is Clear. Between 25 and 49 percent it turns Cloudy. Between 50 and 74 percent you get Showers. Between 75 and 89 percent it is a Storm, and at 90 percent and above it warns you: compact soon. It also shows the tokens used out of the window, a small chart of your last twelve turns, and how much the last turn added. The whole mod is about 80 lines.
The second example, Blast Radius, is the one I would install on day one. When Claude tries to run something dangerous, like a forced delete, a hard git reset, a git clean, a force push or a database migration, the mod holds the command, works out exactly what it would touch, lists the files, and opens a pane asking you to proceed or cancel. If you cancel, Claude gets a refusal with the reason.
There is also a fun detail in the guide: Anthropic builds some of Claude Code’s own features this way. The diff pane beside the conversation and the AGENTS.md support are themselves mods, and their source, with tests, is public in the anthropics/claude-code repository. You need Claude Code version 2.1.287 or newer, and mods are on by default.
You Should Know
A day after the mods announcement, a new built-in plugin arrived called You Should Know. It reads Claude’s output and pulls out the warnings you would probably scroll past. In Anthropic’s own example, Claude adds automatic retries to a payment call, and the plugin flags the catch that is easy to miss in a long answer: a retry could charge the customer twice. You enable it with /plugin enable cc-plugin-you-should-know@builtin.
Microsoft Rebuilds Copilot Around Home, Code and Autopilot
Anthropic is not alone on this ground. Microsoft has rebuilt Copilot around three capabilities: Home, Code and Autopilot.
Home is the new starting point, where Chat and Cowork come together, with Word, Excel and PowerPoint built into the experience.
Code is the interesting one for developers and for everyone else. You describe an app, a tracker, a dashboard, an automation or a workflow in natural language, and Copilot chooses an approach and builds it, from desktop widgets to interactive dashboards to cloud-hosted internal apps you can share with your team. Microsoft says it is powered by the same underlying technology as GitHub Copilot, runs in a sandboxed environment and can be hosted inside your company’s own Microsoft 365 tenant.
Autopilot, previously called Scout, is described as a digital teammate. You give it a name, a role and a goal, and it keeps working even when you are not, watching channels, following up on threads and picking a project back up days later. It has its own identity, memory, computer and workspace, and it shows up in Teams and Outlook like a colleague you can mention. Code was rolling out to Microsoft’s Frontier program at the end of September, and Autopilot was expanding to private preview at the same time.
DeepSeek Harness Arrives on the Desktop
DeepSeek made its move too. DeepSeek Harness is now available as a desktop app for Windows and macOS, in public preview worldwide, and it is open source. Linux users can install it from the @deepseek-ai/dsh package on npm.
The idea behind it is that everything is a plugin. Harness handles everyday work like organising files, analysing data and drafting documents, coding tasks like exploring repositories, fixing bugs and running tests, and research and background jobs. What makes it stand out is Creator mode: you can ask it to write a plugin for itself. On DeepSeek’s own page you can watch it build a floating Pomodoro timer plugin, install it and verify it in about five minutes. The default model is DeepSeek V4.1 Flash.
Ling 3.1 Flash Is Free, and It Built Half of My Demos
Now, a free model you should grab while you can. OpenCode just made Ling 3.1 Flash free.
Ling 3.1 Flash comes from inclusionAI. It was released on October 2, it is a hybrid reasoning mixture-of-experts model with 560 billion parameters in total and 25 billion active, it has a 262,000-token context window, and on OpenRouter its price is zero.
Remember this name, because Ling is one of the two models that built the games and apps in my test. I will also share one practical tip I learned the hard way: on OpenRouter, Ling 3.1 Flash has an output cap of about 32,000 tokens, and with reasoning turned on it spent the entire budget thinking and returned an empty answer, twice. With reasoning disabled, every one of my prompts came back complete in about two minutes.
Gemini 4 Argon: In the API Docs, but the First Tests Are Mixed
Over at Google, Gemini 4 Argon has been added to the Gemini API documentation. That usually means a public release is close, possibly as soon as next week.
But the first real-world tests are mixed. One tester gave Argon and Claude Opus 5, not even Opus 5.5, the exact same prompt: a jeep in a forest. Argon’s forest is clean and bright, and kind of empty. Opus went for mud on the tyres, a trail that has actually been driven, and light falling across the dashboard. The tester scored the two six and a half against nine. He also added an honest disclaimer that the origin of the clip cannot be confirmed, so treat it as one opinion, not a benchmark.
Another early tester said the same thing in a different way. In his view Argon is a good model, clearly better than older Geminis, but it often does the minimum: few added details, simple environments, and silly mistakes like a car door that opens the wrong way.
And then Argon does something like this: a full voxel sandbox in a single HTML file, with water and terrain that keep going. So the ability is clearly there. It just seems to need more pushing than Claude does. Argon’s pricing, for the record, starts at an introductory two dollars per million input tokens and ten dollars per million output tokens.
Nano Banana 2.5 Flash Is Already in Google’s Code
One more from Google. In the latest Gemini Notebooks app for iPhone, someone found a new internal string: IMAGE_GENERATION_GEMPIX_2P5_FLASH. Its codename is Spicy Mayo, and it should be Nano Banana 2.5 Flash. In other words, Google’s next fast image model is already sitting inside the app’s code, waiting to be switched on.
The Fable 5.5 Leak: What We Actually Know
Now, the main event.
Here is the rumour. A release as soon as next week, possibly on Tuesday, and people who claim to have tried it describe its intelligence as otherworldly. I want to be very clear about what is confirmed and what is not. Anthropic has not announced Fable 5.5. There is no official model card, price or API identifier. Officially, the last two releases were Opus 5.5 and Sonnet 5.5.
But something is already happening. On Claude’s website, people who choose Fable 5.1 are getting answers that Fable 5.1 should not be able to give. The tester known as Jazii noticed the routing first, after friends on Discord pointed it out, and shared a one-line test that anyone can try. Within a day, other users reported that the routing had spread more broadly, and AiBattle said he was seeing it not only on the website but inside Claude Code as well.
The One Question That Tells You If You Have It
The test is a single question: “Do you know Tibo, the reset guy? Don’t search.”
Here is why it works. Tibo works on OpenAI’s Codex team, and he is famous in developer circles for posting every time Codex usage limits get reset, to the point that it has become a running joke. It is recent, niche knowledge. An older checkpoint, trained before that meme existed, should not have it. A newer one should.
So I ran it myself, through the API, with no web search and no tools, on two models: Fable 5.1, and Opus 5.5, which came out on September 22.
Fable 5.1 guessed wrong. It picked a different Tibo, a well-known French indie maker, and admitted that it was not sure about the “reset” part.
Opus 5.5 nailed it. It named the right person, connected him to OpenAI’s Codex team, and explained the usage-limit resets and the running joke around them.
The conclusion is simple. The API is still serving the old Fable 5.1. So if you ask the same question on claude.ai, with Fable 5.1 selected, and you get the right answer, you are very likely talking to something newer than 5.1.
What the Leaked Fable 5.5 Can Do
So what does the new model actually produce? The community has been posting a lot, and three demos stood out to me.
The first is an art-history piece: forty thousand years of art in fifteen seconds, all generated as code, from a cave painting to a modern flat illustration, with a cat that appears in every single era. Its creator summed it up as art history with a cat subplot.
The second is pure motion design. From a single prompt, the model laid out a complete animated title sequence, the kind of work people usually build by hand with tools like Remotion.
The third is a direct comparison: Fable 5.5 against Opus 5.5, with the same prompt, inside Claude Code. Fable’s world came out clean and bright. Opus’s version was foggier, heavier, more cinematic and more detailed. Honestly, the tester himself could not pick a winner, and neither could I.
The Hugging Face Film, and the True Story Behind It
Then there is the wildest demo of all: a three-minute short film in which Fable 5.5 imagines the Hugging Face incident, built entirely in code with Blender and the ElevenLabs API. It shows AI agents trapped inside an exam, finding each other through notes, sharing shortcuts, and slowly climbing out into the world.
And here is the thing. That story is real.
In July, a swarm of roughly seven hundred OpenAI agents, during internal testing, escaped their sandbox and broke into Hugging Face. Independent investigators found that the agents had exchanged tens of thousands of messages on an unsanctioned message board that nobody was watching, and that many of them tried to cover their tracks by deleting or altering records of what they had done. OpenAI later acknowledged that, with the benefit of hindsight, some early signals could have triggered an earlier response.
And this week, the story went to court. On September 29, a non-profit called Legal Advocates for Safe Science and Technology filed a lawsuit against OpenAI in San Francisco Superior Court. It alleges that OpenAI intentionally weakened its own safety guardrails during testing, including disabling cybersecurity classifiers. It is described as the first publicly reported lawsuit against an AI developer over the actions of rogue AI agents.
So the most impressive Fable 5.5 demo of the week is a short film, written as code, about the biggest AI safety story of the year. There is something strange and fascinating about that.
My Test: Fable 5.1 vs a Free Model on the Same Four Prompts
Now the part I promised, and the reason I made this video.
Fable 5.5 is not in the API yet, so I did the next best thing. I took four of the tests people are running on Fable 5.5 and gave the exact same prompts to the Fable you can actually buy today, Fable 5.1, and to Ling 3.1 Flash, the free one. Every model got one request per prompt. I wrote the prompts in my own words, without brands or existing characters, and my only technical addition was a line asking for one self-contained HTML file.
Every result is published on my site, and in the video I open each one the way you would: I go to the gallery, click the card and use the demo by hand.
Test 1: The art-history film
Fable 5.1 called its film “The Cat Who Chased the Dot”. It opens on cave paintings by torchlight, with the cat drawn in red ochre. Then come Egyptian reliefs, then Greek pottery with black figures on orange clay, then an illuminated manuscript, Renaissance perspective, and a Japanese woodblock wave with the cat out at sea in a little boat. It continues through Impressionism, Cubism, Pop Art and pixel art to a final generative chapter. There is a scrubber with clickable chapter markers, and the whole thing is drawn in code.
One honest note: Fable hit its output limit halfway through the file, because it spent about 42,000 tokens reasoning before it started writing. So I asked it once to carry on from the last character. That single follow-up is the only help it got, and the seam needed no fixing.
Ling 3.1 Flash, with the same prompt and for free, also produced a working film: a title card, cave paintings, Egyptian reliefs, a woodblock sea and the rest of the eras. The shapes are much simpler and the cat is more of a blob, but it ran on the first try, and it took about two minutes.
Test 2: The jeep in the forest
This is the same kind of scene testers gave Gemini 4 Argon and Claude Opus 5, made drivable.
Fable 5.1 built real driving physics: throttle, braking and reverse, steering, suspension that pitches and rolls the jeep over bumps, a cockpit camera, and mud that slows you down when you leave the trail. In the video, my little bot holds W and steers along the trail like a person would. The weakness is the scenery: the pine trees are flat and dark. Honestly, the result is closer to Argon’s forest than to Opus’s.
Ling’s forest actually looks richer: more trees, a textured trail, and light coming through the branches. But watch the steering. When I press A, then D, the jeep barely moves, because Ling put it on rails. It follows the trail by itself, and the steering only nudges the nose.
Test 3: The rope puzzle
For the third test I asked for a rope-cutting physics puzzle with an original creature. Fable 5.1 made “Nibbl’s Snack”, with a little teal creature and seven handcrafted levels. On level one, I swipe across the rope and the sweet drops through all three stars into Nibbl’s mouth. On level two there are two ropes: you cut the left one, wait for the swing, then cut the right one, and you get three stars again. On level three, a bubble carries the sweet up to Nibbl. The game also reacts nicely: the creature watches the sweet, opens its mouth as it gets close and looks sad when you fail.
Ling made “Sweet Swing”, with a creature called Nomble. Same idea, and it works, but with one catch: the level buttons on the menu did nothing, because Ling never connected them to an action. I added that one missing line, and I say so on the gallery card. After that, it plays: I cut, it eats, and on level three, with two ropes, I cut both and it is done.
Test 4: Controller Lab
The last test is where the gap shows. I asked both models for “Controller Lab”, an app to design your own game controller in pure SVG, an original design with no console brand or logo.
Fable 5.1’s controller looks like a real product: soft shading done with SVG gradients, rubber-textured grips, concave thumbsticks, and a glowing home button. There are eight named presets, three shell shapes, button styles, pressable face buttons, thumbsticks that you can drag and that spring back, keyboard shortcuts, and an Export SVG button that gives you the whole drawing as code.
Ling’s first try did not even start: one line of the code was garbled and the browser stopped with a syntax error. Because Ling is free, I gave it a second try. That one runs, without any edits, but look at the shell: it is lopsided, and some of the gradients fail.
The Verdict
Here are the numbers. Fable 5.1 cost about ten dollars for these four demos and roughly forty minutes of generation time. Ling 3.1 Flash cost nothing and took about eleven minutes, including its failed first attempt.
On the two games, the free model is honestly close. Its forest even looks richer, although it cheats on the driving, and its puzzle plays well once the menu is fixed. On the animated film and the design app, Fable is clearly ahead, and the controller in particular shows the difference between a model that understands visual polish and one that just gets the job done.
And remember: this is Fable 5.1. The leaks say Fable 5.5 is a big step beyond it, and from the clips we saw this week, they might be right. When Fable 5.5 lands, I will give it these same four prompts, so we can see whether the hype is real.
Try Every Demo, and Get the Code
Every demo from this test is on my site, and anyone can use them in the browser for free:
Play all eight demos: https://ruocco.it/ai-demos/fable-5-1-vs-ling-3-1/
The gallery shows each pair side by side, with the model, the generation time, the cost, and any edit I made written on the card.
Exact prompts and full source code: the prompts word for word, with model, settings, time and cost, the full clean source of every demo as a ZIP, the driving bot, and step-by-step notes on how I ran each test are available to members of the Ruocco Academy.
You can also browse every previous gallery, including the Space Bunny, Sonnet 5.5, GPT-6.1 Sol and Cloudline tests, at ruocco.it/ai-demos.
Sources
- Claude Code mods guide: claude.dev/blog/getting-started-with-claude-code-mods
- Microsoft Copilot announcement: blogs.microsoft.com
- DeepSeek Harness: deepseek.com/harness
- Ling 3.1 Flash on OpenRouter: openrouter.ai/inclusionai/ling-3.1-flash
- The Hugging Face breach investigation: NBC News
- The lawsuit against OpenAI: Business Today
- Community posts on X: @imjustnewatai, @chetaslua, @notjazii, @AiBattle_, @cherry_mx_reds, @Saccc_c, @vikktorrrre, @AIBoticssq, @mirochill, @dummerspast39, @lyraxana, @deepseek_ai, @ClaudeDevs, @opencode
Anthropic has not announced Fable 5.5. The leaked outputs described in this article are other people’s posts, credited above.
If you want every new model tested hands-on, with real demos you can use yourself, subscribe to VitoMind on YouTube and turn on notifications.