On June 9, 2026, Anthropic unveiled Claude Fable 5 to the world. Three days later, the US government ordered it taken offline.
That turnaround — from landmark launch to federal shutdown in 72 hours — made headlines on every major tech outlet on the planet. The government issued an export control directive citing “national security authorities,” instructing Anthropic to suspend all access to Fable 5 and its more restricted sibling, Mythos 5.
It was the first time in history that a leading AI company had taken a publicly deployed model offline at the direct order of the federal government.
So what exactly had the government just shut down? And why did it generate this level of alarm in the first place?
Releasing a model this capable comes with risks. Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage.
— Claude (@claudeai) June 9, 2026
Queries on a narrow range of topics will instead receive a response from our next-most-capable model, Opus 4.8. pic.twitter.com/vJ71vCdkjc
Why Fable 5 Is Different From Every AI Model Before It
Until now, the AI industry worked like a toolbox. You had a coding model, a video model, an image model, a music model, a game engine assistant, a data analysis tool. Each one was trained to be excellent in its domain. Each one was largely useless outside it. If you wanted to build something that touched multiple disciplines — say, a game with visuals, sound, and backend logic — you’d stitch together four or five different AI tools, and you’d be the one holding everything together.
Fable 5 made that toolbox obsolete.
It isn’t a generalist that’s decent at many things. It’s a model that performs at specialist level across every major domain — simultaneously. We’re talking about the kind of coding output that senior engineers produce, the kind of visual and UI work that comes from professional designers, the kind of video and animation logic that dedicated media AI tools have spent years optimising for, the kind of computer vision and real-time tracking that used to require custom-trained pipelines, and the kind of creative and strategic thinking that cuts across all of them.
Think about what that actually means in practice. Previously, if you wanted to build a product, you’d use one AI for the code, another for the visuals, another for any video or animation, another for data work, and you’d manually coordinate the output of all of them. The bottleneck was always you — the human glue between a dozen specialised systems. Fable 5 removes that bottleneck entirely. You describe what you want to accomplish, and it figures out which disciplines are required and executes across all of them in a single coherent run.
This is genuinely new. No previous model — not GPT-5.5, not Gemini, not any prior Claude — could credibly claim to be operating at expert level across coding, game development, video creation, visual design, computer vision, data engineering, and autonomous long-horizon work all at once. Fable 5 is the first model where a single person, with a single tool, can do the work that previously required a team of specialists each armed with their own AI.
That’s not an incremental improvement. That’s a different category of technology.
The Numbers Behind Fable 5
To understand why governments are paying attention, you first need to understand what Fable 5 actually scored. Below is a reconstruction of the key benchmarks from Anthropic’s official launch announcement, compared against its previous flagship (Claude Opus 4.8) and OpenAI’s equivalent offering (GPT-5.5).
| Benchmark | What It Measures | Fable 5 / Mythos 5 | Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| SWE-bench Verified | Can the model fix real bugs from real GitHub repos? | 95.0% | 88.6% | ~84% |
| SWE-bench Pro | Harder version: can it solve complex multi-file coding tasks? | 80.3% | 69.2% | 58.6% |
| FrontierCode Diamond | Production-quality code at medium effort | 29.3% | 13.4% | ~6% |
| Terminal-Bench 2.1 | Can it operate a computer terminal autonomously? | 88.0%* | 82.7% | 83.4% |
| GPQA Diamond | Graduate-level science questions (physics, chemistry, biology) | 94.1%* | ~91% | 93.6% |
| Humanity’s Last Exam (no tools) | Extremely hard expert-level questions across all domains | 59.0% | ~47% | 41.4% |
| GDPval-AA (Knowledge Work Elo) | Senior-level analytical reasoning tasks | 1932 | 1890 | ~1870 |
| OSWorld-Verified | Can it control a computer UI like a human? | 85.0% | 83.4% | 78.7% |
| ExploitBench | Cybersecurity exploit discovery and execution | 78.0%* | 40.0% | N/A |
| Legal Reasoning | Complex multi-step legal document analysis | 13.3% | 10.4% | 2.1% |
A few numbers stand out immediately. On SWE-bench Pro — widely considered the most reliable test of real-world coding ability because it uses contamination-resistant tasks from actual public repositories — Fable 5 leads the field by more than 11 points over Opus 4.8 and more than 21 points over GPT-5.5. That’s not a marginal improvement. That’s a different class of capability.
The Humanity’s Last Exam score is equally striking. This benchmark was designed by academics specifically to be unsolvable by current AI. It contains questions so hard that even expert humans often fail them. Fable 5 answers nearly 6 out of 10 correctly without any external tools. Two years ago, the best models were scoring below 10%.
Real-World Examples That Shocked People Online
1) Fable 5 just one-shotted the first open-source MMORPG.
Fable 5 just oneshot the first open-source MMORPG 🤯
— ZYZZ JOBS (@maxpolaczuk) June 12, 2026
Play it on https://t.co/qwlIcpuLCi
Contribute via github: https://t.co/kIpmEpqbm5 pic.twitter.com/C2Dz95IGeR
Someone asked Fable 5 to build a World of Warcraft-style multiplayer online world — in one shot. The result is live at worldofclaudecraft.com and open-source on GitHub. A playable, online, persistent game world with multiple players, character stats, and a 3D environment, generated in a single pass. Building an MMORPG is one of the most complex software projects in existence — it requires real-time multiplayer netcode, a 3D rendering engine, game logic, server infrastructure, and a persistent world state all working together. Studios spend years and millions of dollars on this. Fable 5 did it in one prompt.
2) Fable 5 one-shotted a live traffic tracker for Bangalore — from a single line prompt.
One-shotted a live traffic tracker for Bangalore using Claude Fable and a single line prompt.
— Shubhangi Tiwari (@ShubhangiYaps) June 11, 2026
I’m impressed!
What do you guys think? pic.twitter.com/NWSesyxTrU
Developer Shubhangi Tiwari typed one line: help me build a live traffic tracker for Bangalore. What came back was a fully working, visually polished web app — 24 zones mapped across the city, colour-coded arterial roads, animated dashes showing real-time traffic speed, and hover tooltips with per-road reports. No follow-up prompts, no debugging, no iteration. Previous models given the same instruction would produce either broken code, a static mockup, or a generic template with no real data integration. Fable 5 understood the entire product requirement from a single sentence and shipped it.
3) Fable 5 turned a GitHub profile into a stunning liquid-glass card — with just CSS
Mythos is going crazy with the glass liquid UI with just css https://t.co/x2kuLJwIrd pic.twitter.com/k71lJ9isNX
— Ann Nguyen (@ann_nnng) June 10, 2026
Developer Ann Nguyen asked Mythos to build a GitHub Glass Card tool — something that pulls your GitHub profile and renders it as a downloadable liquid-glass UI card. The output was a fully functional web app: enter any GitHub username, adjust the “card liquidness” slider, add a background, and download your card. The design was striking enough that people thought it used a design tool. It was pure CSS generated by the model. Most AI can write functional code or write pretty code — rarely both at once, and almost never with this level of visual polish from a single prompt.
4) Fable 5 ran autonomously for days — and kept you updated the whole time.
Claude Fable can run autonomously for days.
— Matt Shumer (@mattshumer_) June 11, 2026
This is the single highest-leverage prompt I use to stay on top of it:
"Spin up a persistent HTML page. As you work, append clear, timestamped updates with screenshots/media so I can follow along."
Literally a 10x better experience. pic.twitter.com/TKtsIALzxY
Developer Matt Shumer shared the single most useful prompt he’s found for long autonomous runs: “Spin up a persistent HTML page. As you work, append clear, timestamped updates.” What this unlocked was a model that could work on a task for days — debugging a Chrome DevTools integration, fixing a marketplace download pipeline, building an Unreal Engine showcase — while logging its own progress in plain English in real time. No other model can sustain this. Most drift, lose context, or silently fail after an hour. Fable 5 worked through multi-day tasks and narrated every step, making long autonomous agent runs actually trustworthy for the first time.
5) Fable 5 watched a soccer kick and coached a player — from one goal.
claude fable 5 /goal:
— AA (@measure_plan) June 9, 2026
help me shoot this soccer ball faster https://t.co/8eIKp2Np3x pic.twitter.com/Iac3mvVo57
Developer AA typed: “claude fable 5 /goal: help me shoot this soccer ball faster.” Fable 5 analysed video of a real kick, tracked the ball in real time, calculated its distance to the goal line (11.9 metres), measured its pixel position frame by frame, and delivered specific coaching feedback — all from a single vague instruction with no technical specification whatsoever. This is the kind of task that previously required a computer vision engineer, a custom model pipeline, and days of setup. Fable 5 received a casual one-line goal and built the entire system to achieve it.
6) Fable 5 made an entire video — by itself. Sound effects and all.
Claude Fable 5 made this entire video by itself.
— Nate Herk (@nateherk) June 12, 2026
I gave it a /goal prompt, went to the gym, and came back to this.
Even the sound effects.
Shoutout @HyperFrames_ pic.twitter.com/WCgGI8d6XG
Developer Nate Herk gave Fable 5 a /goal prompt, went to the gym, and came back to a finished 3-minute video — scripted, edited, and complete with sound effects, produced entirely without human input. He didn’t stay to supervise, correct, or prompt it further. He just left. What makes this different from any AI video tool available today is that no single tool does all of this end to end — scripting, visual production, audio design, and final edit are separate pipelines that normally require a human to stitch together. Fable 5 treated the whole thing as one task, made every creative decision autonomously, and delivered a finished product by the time he got back from the gym.
Conclusion: A New Kind of Technology
What makes Fable 5 genuinely new isn’t any single capability. It’s the combination: a model that can produce original science, execute multi-step cyberoperations, compress months of engineering into days, think at a senior-expert level, and sustain coherent judgment across tasks of almost arbitrary length and complexity. That combination, apparently, is enough to make a government reach for the export control lever.
We’re entering a period where the most advanced AI models will increasingly be governed the way nuclear technology or advanced weapons systems are governed — with access determined not just by who can afford to pay, but by who governments trust to use them responsibly. The commercial AI era assumed that the best technology should be available to everyone. That assumption is now being tested.