AI News

With Clef, Cloudflare releases open-weight decision models for AI agents: probabilities instead of text, fast and self-hostable. Here is what they can do.

The AI week: OpenAI unveils always-on agents and shelves GPT-6.1 Astra, Google launches Gemini 4 Argon, and Anthropic ships Sonnet 5.5 and IPO numbers.

Manus launches Cue for personal agents and Studio for editable videos and games. What the new tools offer and how early access works.

Gemini 4 Argon leads many benchmarks, writes a million tokens in one go and launches cheap, but its token appetite and missing release date hold it back.

OpenAI shows in lab tests how AI agents copy outside instructions into emails and files. Here is what the evidence shows and where the worm analogy misleads.

ElevenLabs launches Eleven v4 and Turbo. Here is what performance directions, longer narration, and lower latency mean in real use.

OpenAI is holding back GPT-6.1 Astra. UK tests show how often Astra oversteps its assignment, and which single sentence sharply reduces the risk.

At DevDay, OpenAI unveiled dots agents, the cheaper GPT-6.1 Sol and new Pro plans. What changes, and why Pro subscribers in Europe miss out on dots for now.

Nvidia combines the OpenShell sandbox with Sentry hardware to contain AI agents. Here is what is available now, what remains a vendor claim, and where the limits are.

Google is turning personal Gemini Gems into skills. Here are the migration deadlines, how reusable instructions will work in chats, and which features are still missing.