ChatGPT Voice Now Runs on GPT-6 with Plugins
OpenAI hears the feedback: voice now reaches email, calendar and Slack, backed by GPT-6 Astra, Sol and Luna — on ChatGPT Work, web and mobile.
OpenAI answered its most-requested voice features in a single update. ChatGPT Voice can now invoke plugins for email, calendar and Slack, letting a spoken request turn into a scheduled meeting or a drafted message without leaving the conversation. The assistant is powered by the GPT-6 family — Astra, Sol and Luna — and is available inside ChatGPT Work on both web and mobile, where users can create documents, decks, sites and spreadsheets, or hand off complex multi-step tasks. The move ties voice directly to the workspace tooling OpenAI is pushing hardest this cycle.
Gemini 3.8 Flash TTS Ships Custom Voices
Google DeepMind released two text-to-speech models for creating and deploying custom audio. Gemini 3.8 Flash TTS designs unique voices with distinct accents and characteristics, while Flash-Lite TTS targets efficiency and scale, letting teams choose from created styles or a broad built-in library. The pair extends the Gemini 3.8 line into production-grade voice generation.
Qwen Ships "Intelligence" with Three SOTA Agents
Alibaba's Qwen team introduced Qwen Intelligence, framing personal intelligence as something within everyone's reach. It launches with three state-of-the-art agents, led by a Mobile Planner Agent that plans, decomposes and orchestrates complex tasks and tops MobilePA-Bench across its Business and Memory tracks. The release pushes agentic capability straight onto the phone.
Claude Opus 5.5 Arrives, Cheaper and Faster
Anthropic's newest model is 40% cheaper and over 30% faster than Opus 5, and Higgsfield is among the first to show it off on 3D and creative workloads. The release keeps the Opus line's flagship positioning while trimming the cost of top-tier reasoning and making the model the natural default for demanding agent tasks.
Qwen-Audio-3.1 Completes a Five-Model Stack
ASR, TTS and realtime interaction are fully upgraded, joined by TTS-Next for audio creation and ASR-Next for audio understanding. Together the five models cover understanding, generation, interaction and creation, and ship with significant price cuts across the line.
OpenAI Releases MentalHealthBench
A new open benchmark, built with input from more than 80 mental health clinicians, measures how frontier models perform in realistic conversations — released openly so other researchers can examine it.
Claude Team Makes claude.ai 3x Faster
Using Claude to measure, debug and improve its own performance, the team ran Slack-thread iterations with safety guardrails and shipped roughly 3,000 changes to make claude.ai three times faster in two weeks.
Claude Helps Congo's Ebola Response
CEPI, WHO's Africa office and INRB Kinshasa are using Claude to accelerate the response to an unusual Ebola variant for which no confirmed-effective vaccine yet exists.
TPU Megakernel Pushes Kimi K3 to 709 Tokens/s
The Inferact team open-sourced a single Pallas kernel running all 92 of K3's mixture-of-experts layers, with weight prefetching that reaches across layer boundaries so transfers overlap with compute. It hits 709 tokens/s against 450 on GB200 in low-concurrency decode.
Black Forest Labs Open-Sources FLUX 3 Action
A 7B world-action model reaches SOTA on RoboLab and other leaderboards, with both the backbone and embodiment-specific fine-tunes released as open weights — a step toward broadly available world models for robotics.
Qwen-Image-2.1 Tops Both Image Arenas
Seedance 2.5 Draft Mode Hits 10¢ a Second
Grok 4.7 Climbs the Leaderboard
Claude Code Cloud Sessions Go GA
Work keeps running after you close your laptop; Pro users get $100 and Max users $250 in trial credits.
Runway Lands in DaVinci Resolve
Generate, edit and upscale directly inside the editing timeline without leaving the project.
Cursor Ships Rollouts Monitoring
Writes a monitoring plan, then watches changes as they deploy so regressions are caught before users see them.
Recraft V4.1 Flash Generates in 1.3s
Marketed as the fastest image model on the market, with a Refine step to clean up the details.
vLLM v0.30.0 Lands with 762 Commits
315 contributors ship hybrid-attention hot paths for Kimi K3, DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next.
Copilot Adds GPT-6 Sol and More
Two additional GPT-6 family models become generally available in GitHub Copilot.
NVIDIA Ships Nemotron 3 Diarization
Tracks who spoke when several people talk at once, tuned for voice applications.
GGUF Models Run Directly in Transformers
ggml's Metal kernels enter the Transformers ecosystem, improving local inference efficiency.
Never bet against open-weight models.
Hugging Face Talks Agent Cyberattacks at the UN
The first company to disclose an agent cyberattack calls for more transparency and more open-source AI to empower defenders.
Apple Drops a Qwen3.5-9B Fine-Tune
A new model on Hugging Face that turns long documents into small paragraphs, among other tasks.
DiffusionGemma-Jev Runs on vLLM
Handles yes/no, multiple-choice and scored questions, reading a confidence distribution from each answer slot in a single pass.
NVIDIA Ships SWE-Serve Benchmark
Distilled from 83 SGLang PRs into 53 repo-level tasks, it makes the gap from local checks to full serving paths measurable.
ESA and Mistral Deepen AI Cooperation
The partners will explore secure and trusted European AI applications.
Successful Agents Need Brains, Hands and Files
rauchg distills the winning formula: a model for logic, tools for action, and memories, skills and repos for state.