Models & Research

Version 2.1.170 brings Claude Fable 5 into Claude Code. Just before it, 2.1.169 shipped two things that help more day to day than they sound: a safe mode and a /cd command with no cache loss.

Google Gemini suffered a multi-hour outage today. Error codes 1076 and 1099, no workarounds, over 1,400 users affected. Google confirmed the issue.

OpenAI is opening a platform for external research on the economic effects of AI. Selected researchers get privacy-protected access to data. Sounds like transparency — and it's still a clever move.

Alphabet is raising a record sum for its AI buildout - around 85 billion dollars, with 10 billion from Berkshire Hathaway. What that says about the capital intensity of the race.

Apple finally revealed the new Siri AI at WWDC 2026. Under the hood: Google's Gemini models and Nvidia GPUs. But Europe and China are left out.

A 2-million-token context window, a Deep Think mode, and the ambition to beat Claude and GPT. Gemini 3.5 Pro is supposed to ship in June — but it still isn't here.

Anthropic has hired chemists and is systematically testing how well Claude handles analytical chemistry. First up: reading NMR spectra like a professional.

Tokyo-based Sakana AI has founded a dedicated lab for recursive self-improvement. Just one day after Anthropic warned about exactly this trend.

Apple's WWDC 2026 kicks off Monday. It's Tim Cook's last keynote as CEO — and Apple needs to deliver on AI. Siri is getting Gemini under the hood and becoming a ChatGPT competitor.

China's best-known AI lab used to take pride in running without venture capital. Now DeepSeek is raising money for the first time — with Tencent and battery giant CATL on board.

A Chinese startup releases an open-weight model that takes on GPT-5.5 and Gemini on SWE-Bench. Weights drop in ten days.

Nvidia unveiled Nemotron 3 Ultra, its largest open model: around 500 billion parameters, 5x throughput — reportedly the best open US model, yet still trailing China's top open weights.

Mistral counters the US giants with a complete model family: Large 3 with 675 billion parameters and 256K context, plus nine small models down to the Jetson board.

Google unveiled a tiny board for on-device AI at I/O. The Coral Board runs Gemma 3 270M entirely locally — built for headphones, AR glasses, and smartwatches.

Google just launched Gemini Spark for US users - an AI agent that works around the clock, even when your computer is off. For $99.99 a month.

Europe's biggest AI player is considering its own semiconductors, opening a data center near Paris, and signing deals with Airbus and BMW. Mistral means business.

Leaked renders reveal Apple's new standalone Siri app: a chatbot with document uploads, chat history, and Gemini under the hood. WWDC is around the corner.

Anthropic's new flagship model prioritizes honesty over confidence. Opus 4.8 would rather say 'I don't know' than give you a wrong answer.

Mistral renames Le Chat to Vibe and turns it into a full agent that reads your email, writes reports, refactors code, and ships pull requests. One app for everything.

OpenAI polishes GPT-5.5 Instant for more readable answers but pulls the Canvas panel from its new models. At the same time, o3 and GPT-4.5 are being phased out of ChatGPT.