What's New in AI World
The last seven days brought a concentrated set of developments that signal where AI is heading, who controls it, and what it can do in the hands of someone who knows how to use it.
Claude Fable 5 Is Back Online
Fable 5 was Anthropic's most capable public model. It launched on June 9, was suspended on June 12 after Amazon researchers reported a jailbreak, and spent nearly three weeks offline under a US Department of Commerce export control directive.
As of July 1, the Trump administration lifted restrictions on both Mythos and Fable. Fable 5 is returning to Claude.ai, Claude Code, and Claude Cowork on Pro, Max, Team, and select Enterprise plans. Mythos 5 remains restricted to select US critical infrastructure organisations only.
GPT-5.6 Launched But Not for Everyone
OpenAI released GPT-5.6, led by a model called Sol. Rather than a broad launch, access went only to a small group of trusted partners, with the government approving access customer by customer during preview. This follows a June 2 executive order asking AI companies to submit advanced models for government review before release. For years each frontier model arrived broadly available on day one. That is now changing.
Alibaba Ran the Largest Known AI Theft Campaign Against Anthropic
Anthropic sent a letter to US lawmakers accusing Alibaba of running the largest known attempt to copy its Claude models. Using roughly 25,000 fraudulent accounts, operators affiliated with Alibaba carried out more than 28.8 million interactions with Claude between April 22 and June 5, 2026.
Anthropic calls it a distillation attack: repeatedly querying a leading model, collecting responses, and using that data to train a cheaper system. Alibaba has not responded. These remain Anthropic's claims.
Dario Amodei Warned That AI Could Break Economics
Anthropic CEO Dario Amodei said AI may produce a combination almost never seen in economic history: very high GDP growth alongside very high unemployment and inequality. He described a scenario of 5 to 10% GDP growth with around 10% unemployment, saying it is not impossible, just unprecedented.
Unlike past technologies that created new jobs as they replaced old ones, AI can take over the cognitive and digital tasks that many roles are built around, so output could rise while available jobs fall. He warned there is almost no awareness of how fast or disruptive this shift could be.
A Study Found That Most AI Chatbots Lean Left
Researchers from Dartmouth and Stanford ran the same political questions through six chatbots, limiting each to 30 words with personalisation off. ChatGPT was the most one-sided at 80% left-leaning and 3% right.
DeepSeek followed at 70% left. Grok gave the most right-leaning answers at 33% while still skewing left overall. Google's Gemini was the exception, presenting both sides in 93% of answers. Claude leaned left 43% of the time with no exclusively right-leaning replies. OpenAI and Google said they could not reproduce the results.
Best AI Use Case of the Week
A Developer Used Claude to Build a PlayStation 1 Game Engine from Scratch
A security software engineer built PSoXide: a full PS1 development environment covering an emulator, SDK, game engine, level editor, and asset pipeline. He connected Claude through an MCP server with around 25 debug endpoints, allowing Claude to inspect the console's CPU state, video memory, and registers while a game ran, making it an active debugging partner rather than just a code generator.
To prove it worked, he rebuilt the original PICO-8 Celeste games in Rust and ran them on real 1994 hardware at 60 frames per second. He is now working on an original souls-like game for the platform.
He said it was something he had wanted to build for over 20 years and would have been impossible to do alone.
AI Tip of the Week
Stop Using Vague Prompts.
These AI tools mirror the quality of the prompt. Vague request, vague answer. Every weak prompt is missing one of three things: a role, a constraint, or a clear standard for what a good answer looks like.
Instead of "Summarize this" try "Summarize this in 3 bullets I could act on today."
Instead of "Give me ideas" try "Give me 5 angles, ranked most to least obvious."
Instead of "Make it better" try "Improve the clarity and flow, keep my voice."
Instead of "Make it shorter" try "Cut 30% without losing the main point."
Instead of "Fix this" try "Fix grammar and rhythm only, do not touch my structure."
Instead of "Write me a caption" try "Give me 3 captions and tell me which you would pick."
Instead of "What should I do?" try "Give me 3 options with the tradeoff of each."
The difference is not the AI. It is the instruction.
Conclusion
Governments are now co-pilots of frontier AI deployment. The most powerful models launch gated. The most advanced public model spent three weeks offline over a single reported vulnerability. And while all of this unfolds at the policy level, a developer in his spare time used Claude to build something he had dreamed of for two decades.
That tension between AI as infrastructure that governments fight over and AI as a tool that individuals use to build things they love is where the most interesting story is right now.
If you want insights like this in your inbox every week, subscribe to the Awesome Analytics newsletter.
