Google's Gemini 3.8 Live Beats GPT-Live-1 On Voice At A Fraction Of The Price

Today's AI news: Google DeepMind released Gemini 3.8 Live and 3.8 Live Extended Thinking, voice-first models that reason and run tools in the background while talking, switch between 97 languages mid-sentence, top Artificial Analysis's Speech to Speech Index at 82.6 against GPT-Live-1 Astra's 81.5, and cost 84 cents per hour of input audio against almost $6 for GPT-Live-1. Grok Imagine added in-place text editing on any image in beta. ChatGPT co-inventor Diogo Almeida launched Jev from Typesafe AI, a System One model that outputs only typed decisions with calibrated probabilities, claiming up to 200x the speed of frontier LLMs. Salesforce in Claude opened in beta with 37 pre-built sales skills. And Runway Labs showed a drawing app that animates sketches live with a real-time video model. Chapters: 0:00 Today's AI News 0:31 Gemini 3.8 Live 1:29 Grok Imagine Text Editing 2:00 Jev 2:51 Salesforce In Claude 3:31 Runway Live Drawing Blend Roundup 2026-09-15 https://x.com/GoogleDeepMind/status/2099907440422830269 https://x.com/OfficialLoganK/status/2099909465705447807 https://x.com/imagine/status/2099941522741600502 https://x.com/CompleteSkeptic/status/2099925682726002904 https://x.com/claudeai/status/2099876514330206578 https://x.com/Benioff/status/2099897671305830590 https://x.com/runwayml_labs/status/2099537262698754557 Google's Gemini 3.8 Live thinks and runs tools in the background while it talks, and tops the speech leaderboard. Grok Imagine can now edit the text on any image. A ChatGPT co-inventor emerges from stealth with a model that never generates a word. Salesforce lands inside Claude with thirty-seven sales skills. And Runway animates your sketches as you draw them. Here's today's AI news. Google DeepMind has released Gemini 3.8 Live and 3.8 Live Extended Thinking, its new voice-first models. The Extended Thinking version reasons and speaks at the same time, so it says "let me check that" while it works, and both models run tools and API calls in the background without breaking the conversation. They detect and switch between ninety-seven languages mid-sentence and take live visual input. On Artificial Analysis's Speech to Speech Index, Extended Thinking scores 82.6, edging GPT-Live-1 Astra at 81.5 and Grok Voice Think Fast at 81.3. The cheaper 3.8 Live costs 84 cents per hour of input audio against almost 6 dollars for GPT-Live-1. Available now in the Gemini API and AI Studio. Five days after OpenAI put GPT-Live-1 in the API, Google has undercut it on price and edged it on quality. Voice is the new front line. Grok Imagine can now edit the text on any image. Upload an event invite, a poster, an ad or a screenshot, tell it what the words should say, and it rewrites them in place, matching the font, colour, perspective and lighting so the change doesn't look like a change. It's in beta and xAI is asking for feedback. Legible text has been Imagine's strongest suit since version one, and this takes that from generation into editing. If it works well, it could tempt a lot of users over to Grok Imagine. Diogo Almeida, who co-created RLHF and ChatGPT at OpenAI, has spent two years in stealth on a question: if chat models are superhuman, why haven't they automated the easy work? His answer is Jev, from his company Typesafe AI. It's what he calls a System One model: it never generates text, only typed, structured decisions with calibrated probabilities, trained with a new method called RLCD. That's classification, routing, scoring and extraction, the smart if-statements inside software. He claims that Jev achieves 70 to 500 milliseconds per answer, 4.2 cents per million input tokens, free output tokens, and up to two hundred times faster than frontier LLMs. Currently it's waitlist only. A frontier lab veteran betting against the chatbot is worth watching. Salesforce in Claude is now in open beta, the first piece of the Claudeforce partnership announced last month. It's a plugin that brings your accounts, opportunities and pipeline into Claude with thirty-seven pre-built sales skills, co-engineered by both companies rather than prompts wrapped round the REST API. Prep a call, review deal health, build a pipeline dashboard or send your forecast without leaving the conversation, with Salesforce permissions intact. Marc Benioff's post was two words and a heart: welcome Claudeforce. Salesforce has spent twenty-five years getting people into its app. Now it's putting the CRM where they already are. Runway Labs has shown its first experiment with real-time video models: a drawing app that animates your sketches as you draw them. Scribble a bird and it flaps, sketch a wave and it rolls, with a video model generating every frame live underneath so each stroke becomes motion instantly. It's an experiment built on the real-time model Runway trained on Vera Rubin with time-to-first-frame under a hundred milliseconds. Video generation started as a render you waited for. This is video generation as an input device. Amazing.