Models & Research

The best AI agents achieve only half the performance of PhD experts on complex tasks. A Nature study pumps the brakes on the agent hype - while delivering some surprising nuances.

OpenAI's next model codenamed 'Spud' has finished pre-training. Polymarket gives it 78 percent odds of launching by end of April.

OpenAI releases a specialized cybersecurity model. Their response to Anthropic's Mythos came faster than expected.

Stanford's new AI Index reveals a growing divide between AI experts and the public. Only 33 percent of Americans believe AI will improve their jobs. Meanwhile, China is catching up fast.

On its Q1 earnings call, Goldman Sachs made official what had only been leaked: the bank is working with Anthropic on cyber risks tied to Claude Mythos.

Stanford HAI's ninth annual AI Index lands with a thud: the US-China performance gap has basically vanished, while the biggest labs have quietly stopped publishing almost anything about how their models are built.

Google rolls out interactive 3D models and real-time simulations in Gemini. Physics you can touch, right in your browser.

Z.ai released GLM-5.1 under MIT license: 754 billion parameters, #1 on SWE-Bench Pro, and the ability to work autonomously for eight hours straight. This changes the game.

Starting April 7, Veo 3.1 Fast pricing in the Gemini API drops 14 to 33 percent. Veo 3.1 Lite is already on the table as a cheap alternative — and OpenAI is shutting Sora down this week.

Opus 4.6 and Sonnet 4.6 can now produce up to 300,000 tokens in a single Message Batches API call – unlock with a beta header.

Meta wants to release its next generation of AI models as open source. It would be the first model launch under new AI chief Alexandr Wang — and a clear signal against the company's recent drift toward closed source.

Google releases Gemma 4 under Apache 2.0. Four models from 2B to 31B parameters that claim third place on Arena's leaderboard.

Alibaba releases Qwen 3.6-Plus, a new flagship model that aims to compete with Claude on coding benchmarks and natively supports one million tokens of context.

Alibaba just dropped Qwen 3.5 Omni — a multimodal model that processes 10 hours of audio, speaks 113 languages, and even does voice cloning. And it's fully open source.

French AI company Mistral secures $830 million in debt financing to build its own data center south of Paris. European AI sovereignty doesn't come cheap.

A new import tool makes switching from ChatGPT or Claude to Gemini dead simple. Upload your chat history as a ZIP file and you're done. Google is playing offense.

Google's monthly feature update brings five new additions: free Personal Intelligence, AI music generation, and a new voice model.

A configuration error exposed nearly 3,000 unpublished documents — including details about a new AI model called Claude Mythos that Anthropic says poses 'unprecedented cybersecurity risks'.

Mistral releases a text-to-speech model that speaks nine languages, clones voices from five-second samples, and runs on edge devices. Open source, naturally.

New 'Switching Tools' let you transfer your chat history and personal data from ChatGPT or Claude directly into Gemini.