Models & Research

Unconventional AI just dropped its Un-0 model series. The pitch: run AI on tiny oscillators instead of traditional chips. Sounds wild, but Jeff Bezos put $475 million behind it.

Google has confirmed Gemini 3.5 Pro won't ship in June. It's the second consecutive time a model announced at I/O has missed its target window.

After two weeks of forced downtime, the US government clears Anthropic's most powerful AI model for over 100 institutions. And Fable 5 could be back as early as next week.

OpenAI drops three new models — but only for a handful of partners. Sol is the new flagship, Terra the all-rounder, Luna the budget option. Here's what we know.

Google moves its image model Gemini 3.1 Flash Image (internally 'Nano Banana 2'), alongside Gemini 3 Pro Image, to general availability. The preview variants are being shut down. The fun new bit: you can now feed in a video to generate thumbnails, posters, or infographics from it.

For months, companies threw money at AI tools without measuring results. Now the hangover is here. Uber is capping AI spend, startups are switching to cheaper models, and with OpenAI and Anthropic heading for IPOs, the efficiency question suddenly matters a lot.

After Nobel laureate John Jumper, Google is set to lose two more key Gemini minds to Anthropic, Bloomberg reports — three high-profile exits in six days. The driver: the looming IPO.

Mistral has launched OCR 4 — a document recognition model supporting 170 languages, paragraph-level bounding boxes, and self-hosting options. All for $4 per 1,000 pages.

An immunologist loads data from a shelved experiment into GPT-5 Pro — and within minutes the model delivers the mechanistic explanation his lab had been stuck on for three years.

GLM-5.2 is drawing a clear statement from the field this week: Nathan Lambert calls the open model a 'step change' for open agents. Are freely available models finally on par with Claude and GPT?

Tokyo-based Sakana AI launches Fugu, a multi-agent system that feels like a single model. A 7-billion-parameter conductor decides which expert models get called in.

Anthropic's robot dog experiment enters round two. Claude Opus 4.7 completes all tasks 20 times faster than the fastest human teams — with zero robotics expertise.

OpenAI's life-sciences model is now available as a research preview in ChatGPT, Codex, and the API. It comes with a Life Sciences plugin for Codex that connects researchers to 50+ scientific tools and data sources.

Noam Shazeer, co-inventor of the Transformer and co-lead of Gemini, is leaving Google for OpenAI. The AI talent war just escalated.

Z.ai releases GLM-5.2 — an open-weights coding model with 1 million token context that claims to beat GPT-5.5 on coding benchmarks.

Google unveils its first smart speaker built for Gemini. At $100, AI-powered home automation goes mainstream.

Moonshot AI shipped a 1-trillion-parameter coding model under a modified MIT license — 30 percent fewer reasoning tokens than its predecessor, and a price that stings. The big players, that is.

Google's new open-source model doesn't write text word by word – it guesses 256 tokens at once and sharpens them, up to four times faster than Gemma 4. A fascinating experiment with one clear catch.

Europe's AI champion Mistral AI is eyeing a new mega-round: 3 billion euros at a 20 billion valuation. That's nearly double from September.

The first hands-on takes on Claude Fable 5 are in. Simon Willison talks about that 'big model smell', Ethan Mollick says it leaves every other public model behind, and Andrej Karpathy calls it a real version jump. What's actually behind the hype.