Why Claude Fable 5 was blocked and 15 other wild AI updates this week

The past few days in AI have been unusually intense: the US government stepped in to restrict access to Anthropic’s most powerful model, Elon Musk’s AI–space empire hit historic valuations, Jeff Bezos came back with a new kind of AI company, and a flood of new tools quietly made everyday work feel a lot more automated.

Claude Fable 5 and Mythos 5 suddenly blocked for non‑US users

Anthropic had just launched Claude Mythos 5 and its public variant, Claude Fable 5, when the US government stepped in. Only three days after launch, regulators ordered Anthropic to block access for all non‑US persons, even if they are physically inside the United States.

When you open Claude now, new chats default back to Opus 4.8, and any ongoing Fable 5 sessions end with an error. For many users, the most capable Claude model simply vanished overnight.

According to Anthropic, US authorities were alerted to a method for jailbreaking Fable 5 and extracting sensitive information, including details about security vulnerabilities. Rumors suggest this jailbreak research may have been done by Amazon, which reportedly used prompt sequences to probe Anthropic’s models for security-related information.

The move also follows months of marketing from Anthropic’s CEO, Dario Amodei, who repeatedly described the new model as potentially “too dangerous.” Critics now argue that this framing may have helped trigger a stronger regulatory response than expected. Some users even joked that Anthropic could just rename it “Opus 4.9” and quietly relaunch it.

The backlash has sparked bigger questions about AI geopolitics. Zoho founder Sridhar Vembu called it a sign that “globalization is dead” and argued that countries like India should accelerate their own sovereign AI models. For a deeper dive into the economic and policy implications, see our analysis of the Claude Fable 5 ban and what it means for AI and the economy.

Elon Musk’s AI–space empire and the trillionaire milestone

In parallel, Elon Musk reportedly became the richest person in history and the first trillionaire after a combined SpaceX and xAI entity went public on US stock markets. The listing raised around $75 billion, with investors valuing the combined operation at roughly $2 trillion.

But this is about more than rockets. Musk’s long-term play is to own the full AI stack: chips, compute, power, communications, and even off‑planet infrastructure. The idea is to eventually move data centers into space and deploy up to a million satellites, creating a massive computing and communications layer in orbit.

In this ecosystem, Starlink provides the network, SpaceX provides launch capability, and xAI provides the intelligence. Instead of just building smarter models, Musk is trying to control everything those models depend on.

Jeff Bezos returns with Promethius, an “artificial general engineer”

Jeff Bezos has stepped back into the CEO role for the first time since leaving Amazon in 2021. His new company, Promethius, isn’t another chatbot—it’s an AI system aimed at engineering and manufacturing in the physical world.

Bezos describes it as an “artificial general engineer.” Instead of training only on text and code, Promethius learns from physics simulations and real engineering data. The goal: compress projects that would take 100 engineers 10 years into something 10 engineers can do in one year.

Promethius is designed to help with simulation, prototyping, optimization, and system design before anything reaches the factory floor. In theory, this should accelerate innovation in hardware, robotics, vehicles, and more.

Bezos argues this will create more jobs, not fewer, predicting long-term labor scarcity and higher living standards. But the claim sits awkwardly next to Amazon’s recent layoffs and heavy automation push, so it remains to be seen whether this is an economic forecast or a sales pitch for his new AI vision.

Moonshot AI’s Kimi Work: 300 local agents on your laptop

Chinese company Moonshot AI launched Kimi Work, a desktop AI agent platform that can run up to 300 agents in parallel—directly on your machine. Unlike most AI agents that send your data to remote servers, Kimi Work is designed to keep your files and workflows local.

Once installed, Kimi Work can:

• Summarize PDFs and documents in your local folders
• Generate sales or financial reports from your own files
• Research online, pull the latest market or trend data, and extract key insights
• Turn all of that into polished PowerPoint decks, web dashboards, or spreadsheets

Paired with Web Bridge, a browser extension, Kimi can also navigate websites and interact with them directly—filling forms, collecting data, and comparing products in your browser.

The tool is especially tuned for finance teams, with built‑in access to sources like Yahoo Finance, the World Bank, and Binance. One standout feature is its ability to deploy sub‑agents, coordinate their research, and automatically assemble the results into a clean dashboard or presentation.

For example, you can ask it to:

• Analyze Germany’s chances at the 2026 World Cup
• Pull historical performance data and current squad news
• Build an interactive prediction dashboard
• Create a five‑slide investment-style summary deck

Kimi Work plans the task, spins up multiple agents, browses the web, builds the charts, and designs the slides—then hands you a working web page and a ready‑to‑present slide deck.

Kimi K 2.7 Code: open-source, agentic coding on your machine

Alongside the desktop app, Moonshot released Kimi K 2.7 Code, its latest open-source coding model. It’s designed for long, multi‑step tasks rather than quick one‑off answers.

Compared to the previous 2.6 version, K 2.7:

• Writes stronger, more reliable code
• Handles multi‑step, long‑running workflows better
• Uses about 30% fewer reasoning tokens
• Runs up to six times faster on similar tasks

To run it, you paste a single command from the Kimi Code page into your terminal, sign in, and choose a permission mode. “Yolo” mode lets it act with full autonomy, while “swarm” enables multiple agents to collaborate on a task.

In one test, a single prompt asked Kimi K 2.7 to write a 4,000–5,000 word, data‑driven blog post comparing three anticipated tech IPOs—SpaceX, Anthropic, and OpenAI. The model independently browsed sources like Reuters, TechCrunch, and CNBC, cross‑checked facts, and generated a full article with an executive summary, company breakdowns, comparison tables, and image placeholders.

Independent users have also benchmarked K 2.7 against K 2.6, GPT‑5.5, and Claude Opus 4.8 on visual reasoning tasks like rendering a realistic water wave. K 2.7 produced the most physically accurate result, highlighting how quickly open models are catching up in specialized domains.

Apple’s rebuilt Siri and on‑device AI

Apple quietly turned Siri into a very different kind of assistant. Instead of just answering questions, the new Siri can see what’s on your screen, understand your context, and take actions across your apps.

In one demo, Siri:

• Recognized a photo taken on the Santa Cruz coast
• Identified it as Natural Bridges State Beach
• Dug a friend’s address out of an old text message
• Built a route in Maps that stopped at both locations

Siri’s voice has been rebuilt to sound more natural and expressive, and you can customize its expressiveness and speaking pace. Apple also introduced a dedicated Siri app, which acts as a central hub for all your AI conversations across iPhone, iPad, and Mac.

Apple’s AI models and the Gemini partnership

Under the hood, Apple isn’t doing this alone. It has entered a deep collaboration with Google, using Gemini technology to power parts of its AI stack. Reports suggest Apple is paying Google around $1 billion a year for this integration—less than the estimated $1.5 billion it would have cost to rely on Anthropic’s models.

However, Apple isn’t just slapping Gemini onto iOS. The company has built its own Apple Foundation Models that run on‑device and on Apple’s private cloud infrastructure. Apple emphasizes that requests are processed without storing personal data, and external experts can audit the system to verify privacy claims. For Apple, trust and privacy are the main selling points of its AI strategy.

AI-powered photo editing with spatial reframing

Apple also brought generative AI into its Photos app. You can now:

• Clean up distractions in the background
• Extend a photo beyond its original frame
• Use “spatial reframing” to change the perspective as if you had moved the camera

Spatial reframing uses on‑device spatial models to let you drag the frame around and see the perspective shift in real time. A generative model then fills in missing areas around the edges. In practice, it means you can fix awkward compositions or bad framing after the fact, even if your friend took a less‑than‑perfect shot.

Google’s Gemini 3.5 live translate: real-time conversations across 70+ languages

Google rolled out Gemini 3.5 live translate, an audio model that lets two people speak different languages and still have a natural, back‑and‑forth conversation in real time.

Here’s how it works:

• You open the Google Translate app and enable live translation
• Person A speaks in their language
• A few seconds later, Person B hears the message in their own language
• Person B replies, and Person A hears it in theirs

Unlike older translation apps that forced you to pause after every sentence, Gemini 3.5 keeps the conversation flowing. It also preserves the speaker’s intonation, pacing, and pitch, so the translated voice sounds more like a human than a flat robot.

Live translate is already available in the Google Translate app and is coming to Google Meet for video calls. For travelers, remote teams, and global businesses, this is one of the most practical AI features Google has shipped so far.

ChatGPT gets built-in interactive charts

OpenAI added a surprisingly powerful feature to ChatGPT: native, interactive charts from a single sentence.

If you ask, “What are the top 10 countries by GDP?” ChatGPT now returns not just a list, but a full bar chart. Hovering over each bar shows the exact numbers—for example, the US at $30.5 trillion and Germany at $4.9 trillion. You can then ask for a line chart of GDP growth over the last 10 years, or a donut chart of energy sources in those countries, and it will generate those visualizations on the fly.

All of this works on mobile as well. With around 900 million weekly users, ChatGPT just turned into a lightweight data visualization tool that fits in your pocket. For many people building reports or client presentations, this reduces the need to learn dedicated BI tools like Tableau or Power BI—if you can ask the right question, you can get a usable chart.

Canva moves inside ChatGPT

Canva has integrated more deeply with ChatGPT, letting you turn AI‑generated images into fully editable Canva designs without leaving the chat.

The flow is simple:

• Describe the image you want in ChatGPT
• Mention that you want it as an editable Canva design
• Open the generated design in Canva and tweak it as needed

From there, you can adjust layouts, swap text, change colors, and export in your usual formats. It’s a smooth bridge between text‑based ideation and full design work.

Figma can now “copy” any live website into editable designs

Figma, the go‑to tool for interface designers, launched a feature that turns any live website into an editable Figma file in one click.

Previously, if you saw a site you loved and wanted to use it as a starting point, you had to manually recreate every section. Now you can:

• Open the site in your browser
• Click the Figma extension and hit “capture”
• Paste into Figma, where the site is reconstructed as real components

Once imported, you can move elements, edit text, swap images, and change colors as if you had designed it yourself. It dramatically shortens the time from inspiration to working prototype.

Notebook LM grows into a research agent

Google’s Notebook LM, originally a research tool that only worked with documents you uploaded, just became much more agentic.

New capabilities include:

• Finding and importing relevant sources on its own
• Generating diagrams directly from chat
• Creating structured artifacts like spreadsheets from your conversation
• Suggesting follow‑up questions and deeper lines of inquiry

Instead of being a static summarizer, Notebook LM now behaves more like a research assistant that can discover sources, analyze them, and hand you ready‑to‑use outputs.

OpenAI Codex: smarter usage resets

OpenAI’s Codex, its coding assistant, has long had usage caps that forced you to wait for a timed reset. The problem was that the reset happened on a fixed schedule, whether you needed it at that moment or not.

Now, resets are more flexible. You can save your reset and trigger it exactly when you hit your usage limit, instead of wasting it when you’re idle. It’s a small change, but for developers in the middle of a build, it removes a lot of friction.

Replit “skills” and custom instructions

Replit’s AI, which can generate full apps from natural language, used to have one big annoyance: you had to re‑explain your preferences for every new project. Brand colors, layout rules, tone, and other details had to be repeated each time.

The new “skills and custom instructions” feature solves this. You can now:

• Create a reusable “skill” that encodes your rules and preferences
• Upload a file with your full brand guidelines
• Apply these skills automatically to any new project

Replit also offers a library of pre‑built skills, including ones modeled on popular tools and workflows. Instead of starting from zero, you teach the AI your style once and it carries that knowledge into every future build.

Hermes agents: scheduled jobs on autopilot

Hermes, a popular free AI agent platform, added the ability to run jobs automatically on a schedule. You no longer have to manually trigger your agents every day.

To set it up, you:

• Pick a “blueprint” such as a morning brief, news digest, or habit check‑in
• Answer a few configuration questions
• Set the schedule and let Hermes run

From there, Hermes can send you daily summaries, track habits, or monitor specific topics, all without further input. It’s a simple but powerful step toward fully automated personal workflows.

Claude + Hyperframes: turning answers into instant videos

Claude can now turn its own text answers into explainer videos directly in the chat using Hyperframes, a tool from HeyGen.

After enabling Hyperframes under Claude’s connectors, you ask a normal question—say, for Tesla projections in 2026. Instead of reading a long text output, you can then ask Claude to convert the answer into a video. Hyperframes automatically scripts, structures, and renders a short explainer, complete with narration and visuals.

You can even tweak the speaking style. For learners and teams who absorb information better visually, this turns dense analysis into something much easier to watch and share. If you’re curious how Claude’s latest models behave under heavy workloads, you might also like our week‑long hands‑on test of Anthropic’s Fable 5.

What this week’s chaos tells us about AI’s direction

Put together, these updates paint a clear picture of where AI is heading:

• Regulation is catching up. The Fable 5 restrictions show that governments are willing to intervene directly when they believe models pose security risks.
• Infrastructure is the new battleground. Musk’s space‑compute vision and Apple’s private cloud models highlight a shift from just building models to owning the full stack.
• Agents are becoming normal tools. From Kimi Work’s 300‑agent swarms to Hermes’s scheduled jobs, multi‑agent automation is moving onto everyday laptops.
• AI is embedding into everything. Siri, Google Translate, Photos, Figma, Canva, Replit, and Notebook LM are all weaving AI deeper into tools people already use daily.

For users and builders, the takeaway is simple: the frontier is no longer just about chatting with smarter models. It’s about agents that can see your screen, browse the web, coordinate with each other, and quietly run in the background—while governments, tech giants, and open‑source communities wrestle over who controls them.