Claude Fable 5 is the most capable model released. It comes in at #1 on the Artificial Analysis Intelligence Index, ahead of OpenAI's GPT-5.5 and a clear step above Opus 4.8. It’s also one hell of a coder. I personally got a few days of coding with it before it went dark, and the jump from Opus 4.8 is real. Running it costs about twice what Opus 4.8 does; that’s going to be costly and total overkill on simple tasks, but for complex ones it might actually turn out cheaper as it can make way more progress in the same number of prompts.

Fable shipped with some unusual guardrails. When asked about cybersecurity, biology, or chemistry it either declines, or routes you to Opus 4.8 instead. Some people say it refuses to answer basic things which are only vaguely related to these topics. The fully unrestricted version, Claude Mythos 5, goes only to vetted partners.

The US government had it switched off. On June 12, citing national security, authorities issued an export-control directive to suspend all access to Fable 5 and Mythos 5 "by any foreign national," after claiming someone had found a way to jailbreak the model. Because Anthropic couldn't reliably separate foreign nationals from everyone else in real time, it pulled both models for every user worldwide within hours. Anthropic calls it a "narrow potential jailbreak" and frames it as a temporary suspension, but with no restore date yet.

The unrestricted version now has a national-security job. Anthropic reportedly embedded engineers inside the NSA to run Mythos for offensive cyber operations against the likes of China and Iran. Pair that with research showing Mythos can weaponize a freshly disclosed bug within hours, and you can see why the most powerful model is being treated like a munition, rather than a product.

Dario Amodei is asking Washington to regulate him. Dario Amodei published Policy on the AI Exponential, arguing (using a Lord of the Rings metaphor!) that democratic institutions move too slow while AI capability now compounds in weeks. He’s asking for mandatory third-party testing of the most powerful models before release, with the gov being able to block or revoke them. Yes, a frontier lab calling for a regulator that could ground its own models.

Anthropic launching and then having their AI withdrawn on national security grounds is the loudest signal that they actually have the best model in the world.

The export control thing has the logic of weapons-grade technology, and is a new flavour of what the US already runs on AI chips to China, now aimed at the model itself.

Anthropic gating its models and, in the same breath, asking the government for the power to gate everyone's? Could be responsible, could be self-serve, could be both.

OpenAI and Anthropic both filed to go public within a week of each other. Anthropic confidentially filed a draft S-1 for a possible fall listing, and OpenAI filed days later, with Sam Altman telling staff a listing could come "within the next year."

SpaceX priced the biggest IPO in history, and the pitch is AI infrastructure. It went out at a $1.77T valuation, raising about $75B. The bull case is about compute: Google is reportedly paying SpaceX $920M a month for capacity. SpaceX is also talking a big game about solar-powered data-center satellites.

Apple finally shipped a real Siri, built partly on Google's Gemini. At WWDC, Apple rebuilt its assistant as Siri AI: it holds real conversations, reads what's on your screen, searches your messages and photos, and takes actions across apps.

Google gave NotebookLM a brain upgrade and its own computer. NotebookLM's chat moved from simple retrieval to a real agent running on Gemini 3.5, and each notebook now gets a sandboxed cloud computer that writes and runs code, builds charts, and does exports. It's quietly turning into a research workstation.

Google's Gemma 4 12B is a capable multimodal model that runs on a laptop. The new open model takes image and audio input, runs on a machine with 16GB of RAM, and is claimed to be nearly as good as the much larger 26B version.

Google shipped a real-time speech translator across 70+ languages. Gemini 3.5 Live Translate does continuous speech-to-speech translation that keeps your tone and detects language switches mid-sentence, live now in Google Translate and coming to Meet.

Frontier agents still flunk real professional work, and the harness matters as much as the model. UC Berkeley's Agents' Last Exam found top agents struggling on realistic workflows, with the harness around the model nearly as important as the model itself (notably, GPT-5.5 beat Claude Fable 5 on this one).