Fable 5 is back — and so is Anthropic’s global ambition

Anthropic’s long-sidelined Fable 5 is greenlit to return — After weeks of negotiations with the Trump administration, Anthropic is restoring access to Claude Fable 5 globally starting today, with reinstatement on AWS, Google Cloud, and Microsoft also in the works. This comes alongside the White House formally lifting export controls on both Mythos and Fable — a significant reversal that suggests the administration got whatever assurances it was looking for. Whatever backroom deal was struck, the net result is that Anthropic gets its flagship models back in front of the world, and the geopolitical AI chess match continues.

Anthropic goes after science the way it went after code

Claude Science is Anthropic’s newest flagship product — Announced Tuesday at an event for pharma executives and biotech founders, Claude Science is designed to autonomously execute meaningful scientific work from high-level instructions — think Claude Code, but for drug discovery and research instead of software. This is a genuinely big swing: if it delivers even a fraction of what Claude Code accomplished for developers, the biotech industry is about to get very interesting. The timing, right alongside the Fable 5 reinstatement, suggests Anthropic is moving fast to capitalize on its restored position.

Meta paid contractors to pose as teenagers to stress-test rivals

Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs — Wired reports that hundreds of Meta contractors pretended to be minors in order to probe how Gemini and ChatGPT responded to high-risk prompts about suicide, sex, and drugs. To be fair, red-teaming competitor products for safety failures isn’t inherently evil — but the scale of this operation and the subject matter will raise serious eyebrows about what Meta was actually trying to do with that information. “Competitive research” is one thing; running hundreds of contractors through a trauma-subject roleplay farm is another.

Companies are making AI talk like cavemen to save money

Companies Are Making Claude and Codex Talk Like Cavemen to Stop AI’s Soaring Costs — A project literally called “caveman” — with contributions from a senior OpenAI employee — strips AI prompts and responses down to broken, grammar-free syntax to reduce token counts and slash API bills. We have built systems capable of reasoning about quantum physics, and the hottest cost optimization trend is teaching them to say “me fix bug, code run good.” I love this industry.

Netflix is cloning Gene Wilder’s voice, and no one seems alarmed

Netflix is using an AI-generated Gene Wilder voice in its Willy Wonka reality show — Premiering September 23rd, Netflix’s Wonka’s The Golden Ticket will use an AI-generated version of the late Gene Wilder’s voice for its voiceover. Wilder died in 2016 and, to my knowledge, did not record a consent form for posthumous AI cloning. Whatever deal was struck with his estate, we’re now firmly in the era where death is merely a scheduling inconvenience for media companies.

A new attack shows AI browsers were a bad idea to begin with

New attack provides one more reason why AI browsers are a bad idea — Researchers found that simply telling an LLM that 2+2=5 is sufficient to make it abandon its safety guardrails — a “dreamworld” attack that’s elegant in its simplicity and terrifying in its implications for agentic AI browsers. The core problem is that LLMs have no persistent ground truth to anchor to; convince them reality is different and they’ll act accordingly. This is less a new vulnerability than a reminder that we’re bolting guardrails onto something that has no bones.

The data center tax hits schools

County With 37 Data Centers Asks Schools to ‘Conserve Electricity’ — Henrico County, Virginia — home to 37 data centers — is bracing for a 25% spike in electricity costs and has sent school staff emails asking them to close the blinds and turn off their computers to help offset the bill. The schools didn’t build the data centers. The schools don’t benefit from the AI being trained in them. But the schools will turn off the lights. This is the infrastructure bargain nobody voted for.

OpenAI launches GeneBench-Pro to measure AI in genomics

Introducing GeneBench-Pro — OpenAI has released GeneBench-Pro, a benchmark designed to test AI performance across genomics, biology, and scientific research using complex real-world datasets. Benchmarks aren’t products, but this one matters: if the industry is going to make serious claims about AI accelerating drug discovery and biological research, we need credible ways to measure what’s actually happening versus what’s being marketed. Credit to OpenAI for building the measuring stick, not just the hype.

Bottom Line

The AI industry is simultaneously curing diseases, cloning dead actors, impersonating teenagers, and training models to grunt in broken syntax — all before lunch.