This week in AI was one for the history books. Between jaw-dropping demos, unexpected leaks, a mysterious stealth model with a million-token context, and Meta’s boldest hardware play yet, there was no shortage of headlines. But one story rises above all of them — and it’s the kind of breakthrough that changes how we think about the role of artificial intelligence in scientific discovery.
Claude, Anthropic’s flagship AI model, found something hidden in the DNA of viruses that no human scientist had ever noticed. Anthropic took that finding into a real biology lab in the Bay Area and confirmed it was real. They named the discovery ART — Array-associated Reverse Transcriptases — a novel enzyme system with CRISPR-like repeating structures that nobody had seen before.
Let’s break down everything that happened this week, from the most impressive AI demos to the most meaningful scientific milestone.
The $1,074 Island — Built Entirely by Claude Opus 5.5
Let’s start with the demos, because this one is genuinely insane.
Dan Green Higgs spent $1,074.40 of Claude Opus 5.5 tokens on a single procedurally generated world. Not credits, not API calls — just pure model output, token by token. The result is something you have to see to believe: a fully realized tropical island where birds fly over turquoise water and dive into the ocean, wooden signs and lanterns sway in the wind, and you can walk into every house, take a boat out to sea, and even go underwater to explore the seabed. Every pixel, every animation, every detail was generated by the model from prompts alone.
This is what a frontier model can do when you stop worrying about the bill. Most of us will never spend that much on a single project — and honestly, most projects don’t need it. But the sheer fact that this is possible from a chat interface tells you where the ceiling is right now. A year ago, generating a single coherent 3D scene from text was research-grade work. Today, one model can generate an entire explorable world in a single session, given enough tokens and the right prompting.
GPT-6 Sol Answered with a Volcanic Rover
OpenAI didn’t stay silent. Chetan Anakola built something equally impressive with GPT-6 Sol: a fully playable 3D world where you drive a six-wheel rover around a reflective volcanic lake toward an abandoned observatory. The sky shifts dynamically from dusk to an aurora-lit night while you drive, and the entire lighting environment changes with it. The rover physics, the reflections on the water, the atmospheric transitions — all generated by the model. And it runs in a browser tab, no GPU required on the client side.
Higgsfield went one step further and put the two models head-to-head inside Unreal Engine through their own MCP (Model Context Protocol). Opus 5.5 on top, GPT-6 Sol on the bottom. Same prompt, same engine, same idea — an underwater scene with a submarine exploring a submerged wreck. Both submarines move, both scenes are full of wrecks, ruins, and fish, and both look like real game prototypes. A year ago, neither of these was possible from a chat window.
GPT-6 Astra Minor — Leaked in OpenAI’s Own Help Center
A model called GPT-6 Astra Minor appeared inside OpenAI’s official help center documentation. AI Battle caught it before OpenAI noticed and removed it. The leaked table listed Astra Minor as a supported mainline model, placed right next to GPT-6 Sol, GPT-6 Luna, and the full Astra. It keeps the standard safeguards under what OpenAI calls Daybreak Blue — the same access level as Astra.
What we don’t know: benchmarks, pricing, context window size, or release date. Nothing. My best guess is that “Minor” refers to a smaller, more cost-efficient version of Astra that fills the gap between Sol (the capable mid-tier model) and the full Astra (the flagship). What’s certain is that this wasn’t a random config string — it sat in the official documentation next to models you can use today.
Space Bunny Alpha — The Stealth Model
A mystery model called Space Bunny Alpha appeared on Open Router and Open Code with no clear provenance. The provider is anonymous. The price is free. The context window is 1 million tokens. It accepts text, images, and video input. AI Battle asked it in Chinese what model it’s based on, and the model’s own reasoning gave it away: “We should state that we are an AI assistant made by MiniMax.”
Then Cheaty found MiniMax M3.1 in MiniMax’s official model catalog, staged but not announced. The strongest clue? The text tokenizer. On six fingerprint prompts, Space Bunny counts exactly the same tokens as the official MiniMax M3 vocabulary. DeepSeek and Kimi don’t match at all.
Someone even built a whole investigation website — spacebunnyalpha.com — showing field notes: 1M context, 500K+ output tokens, ~87 tokens/sec, and a 24/24 tokenizer match with MiniMax. The SVG lab lets you test it with five prompts — Mona Lisa, a pelican on a bicycle, a raccoon riding a dog steering a jet ski — all drawn entirely in code.
Claude Code Gets Real Updates
Claude Code received two substantial updates this week. Search Projects splits work into threads, runs them as parallel Claude sessions, passes context between them, and keeps going when you leave. Claude Sessions are officially out of research preview — you start a task, close your laptop, and Claude Code keeps working on Anthropic’s servers. Existing subscribers get a one-time credit: $100 on Pro, $250 on Max, claimable with /claim credit before October 7th.
Kimi K3.1, MiMo V3’s HySparse2, ChatGPT Voice, Gemini 3.8 TTS
Kimi K3.1 is getting close — a new model ID with 1M context was found in Kimi’s live configuration. MiMo V3 introduced HySparse2 architecture, needing ~5x less prefill compute and a KV cache 4.5x smaller at 1M tokens. OpenAI upgraded ChatGPT Voice to work with plugins and inside ChatGPT Work. Google released Gemini 3.8 Flash TTS — their most expressive audio models yet, with 2,000+ ready-made voices.
Meta’s 100-Gram VR Glasses
Meta unveiled VR glasses weighing about 100 grams — the weight of a deck of cards. 5K micro OLED displays, the first IMAX-enhanced VR device, a Meta AI agent built in, no controllers needed. Ships spring 2027 for $1,299.
The Big One: Claude Discovers a Novel Enzyme System
Now we arrive at the story that matters most.
Anthropic published a page titled “Claude discovers a novel enzyme system with CRISPR-like repeats” — and simultaneously announced the launch of their own dedicated biology lab. This is not a press release about a future capability. This is a documented account of an AI making an original biological discovery.
Many of the biggest discoveries in biology started with someone noticing something odd in nature. CRISPR itself was first spotted as a strange repeating sequence in bacterial DNA that nobody understood for years. Now, Claude autonomously discovered a novel enzyme system. The human scientists only wrote the initial prompt and did the laboratory confirmation. Everything in between — the search, the analysis, the hypothesis formation, the moment of insight — was Claude.
Feng Zhang, one of the pioneers of CRISPR gene editing at MIT and the Broad Institute, personally reviewed the discovery. He calls it “an exciting example of how AI agents can contribute to biological discovery.”
The numbers: roughly 950 Claude agents searching for 21 hours, consuming 210 million tokens. They gathered more than 200,000 reverse transcriptases, identified 3,500 new candidate systems, and narrowed them down to the 20 most compelling, each with a detailed report. Anthropic estimates this would take an expert human weeks to months.
And then this moment: while reading raw DNA sequences next to one unusual enzyme, the Claude agent wrote: “This is spectacular. I can see by eye a tandem repeat array. That’s a CRISPR-like repeat array.” It then counted the repeats, measured their spacing, compared the layout against known systems, and searched the literature. Only when confident, it filed a report.
Anthropic’s scientists took it into their lab in the Bay Area. The discovery was real. They named it ART — Array-associated Reverse Transcriptases. Three parts: the enzyme, a partner gene, and a long array of repeats. Early experiments show the array is expressed as distinct short RNAs — structurally similar to the guide RNAs that make CRISPR programmable.
Nobody knows yet what ART does. There is no evidence it can edit genes like CRISPR. But an AI searched a massive dataset, noticed something every human researcher had missed, formed a testable hypothesis, and handed it to scientists who confirmed it in a physical laboratory. That has never happened before at this scale.
This is only Anthropic’s first program in their new biology lab. Claude didn’t cure a disease today, but this is the first real glimpse of AI starting biological discoveries by itself.
From a $1,074 procedurally generated island to a leaked GPT-6 model, from a mysterious free model that might be MiniMax M3.1 to a 100-gram VR headset — and finally, to an AI making a genuine biological discovery that passed laboratory validation. Which one surprised you most? Let me know in the comments.
Originally published on ruocco.it