Amazon launched Nova Sonic speech-to-speech AI for human-like interactions

—Outperforms OpenAI's voice models with ~ 80% less cost
—4.2% word error rate across languages
— 46.7% better accuracy than GPT-4o for noisy environments
—On Amazon Bedrock

Amazon also dropped an upgraded Nova Reel 1.1 video model

—Delivers improved quality, style consistency
—Extends generations to 2 min via automated and manual, shot-by-shot modes
—Also available on Amazon Bedrock

Google made headlines by making Deep Research available on Gemini 2.5 Pro Exp

The move enables Gemini to create superior research reports over rivals

Also includes new audio overviews to turn reports into podcast-like conversations!

Nvidia released Nemotron-Ultra, a 253B parameter reasoning AI

—Surpasses DeepSeek R1, Llama 4 Behemoth, and Maverick across benchmarks
—Includes a reasoning on/off toggle
—Open-source with model code, weights, and post-training data on Hugging Face

Deep Cogito emerged from stealth with a mission to build "general superintelligence"

The startup also launched Cogito v1 Preview, a family of open-source models that it claims beats the best available open-source models of the same size

Stanford students' research discussion forum AlphaXiv introduced Deep Research for arXiv

The tool compiles literature reviews from trending papers, turning hours of research work into mere seconds of natural language search

Together AI dropped DeepCoder-14B, a new reasoning-capable coding AI

—Matches o3-mini (low) and o1 on competition-level coding tasks
—Also generalizes well to math, scoring 73.8% on AIME
—Fully open-sourced with dataset, code, and training recipe

OpenAI released GPT-4.1, 4.1 Mini, and the ultra-fast 4.1 Nano—all designed for devs

— Each model beats GPT-4o and 4o mini on dev tasks
— 1M token context windows
— GPT-4.1 scored 55% on SWE-Bench Verified
— Starting at $0.10/0.40 per million I/O tokens

Google released DolphinGemma, an AI that can generate dolphin vocalizations

The system analyzes dolphin whistles and sounds to identify patterns and predict subsequent sounds!

Set to be open-sourced soon!

Hugging Face acquired Pollen Robotics, a French startup building open-source humanoids

Pollen is already selling Reachy 2, an open and VR-compatible humanoid for research, education, and embodied AI

A big move from HF in open robotics!

TikTok parent ByteDance just dropped Seaweed, a hyper-efficient 7B-param video AI

—Supports text-to-video, image-to-video, and audio-driven synthesis
—Clips up to 20s
—Matches or outperforms larger models like Sora, Kling 1.6, and Veo

Ex-OpenAI chief scientist Ilya Sutskever's Safe Superintelligence (SSI) has reportedly raised $2B at a $32B valuation

This comes as SSI continues to climb a "different mountain" for developing advanced, superintelligent AI

LM Arena launched a new Search Arena Leaderboard, designed to evaluate LLMs on search tasks

Currently, the top-ranked LLMs on the leaderboard are Google’s Gemini-2.5-Pro and Perplexity’s Sonar lineup!

OpenAI dropped the new o3 and o4-mini reasoner models

o3 pushes SOTA performance across coding, math, science, and multimodality, while o4-mini offers fast, cost-efficient performance

Both have agentic access to ChatGPT tools and can "think with images"

Anthropic added a Research feature in Claude with Google Workspace integration

Research will perform searches across the web and users’ connected work data

This data will also include users' emails, calendars, and docs, thanks to the Workspace link

Cohere released Embed 4, a SOTA multimodal embedding model to add frontier search and retrieval capabilities to AI apps

—128K-token context window
—Supports 100+ languages
—Optimized for data from regulated industries
—Up to 83% savings on storage costs

Microsft added computer use capabilities to its Copilot Studio

The move will allow enterprises to build and deploy agents that can take UI actions on desktop and web apps!

A big move!

Microsoft also started rolling out Copilot Vision in its Edge browser

It will read what's on screen to summarize aloud, working as a real-time collaborator/assistant when browsing the internet.

Best part: it's free—and opt-in (not active by default)!

China's Kling AI released two new models: KLING 2.0 Master for video generation and KOLORS 2.0 for images

Both come with improved prompt adherence, with KLING 2.0 standing out when dealing with prompts with sequential actions and complex motions

Elon Musk's xAI started rolling out a memory feature into its Grok assistant (in beta)

Just like ChatGPT, Grok will reference past chats to provide personalized answers.

There's also a dedicated "forget" button to exclude specific chats from its memory