Cur8 — Thursday, September 24, 2026

Gemini 3.8 TTS models launched: 2,000+ voices, custom voice creation from 30 seconds of audio, and conversational character scripts. Generating 1 minute of audio costs $0.0274 using Gemini Flash TTS.

AWS Bedrock now enables private, cost-effective AI coding agents using open-weight models like Moonshot Kimi K3, GPT-OSS 120B, and NVIDIA Nemotron 3 Super 120B. OpenCode, a new open-source tool, leverages this with multi-model workflows; fine-tuned models outperformed GPT-4o by 35% while reducing costs.

Altman urged UN action on AI governance, emphasizing potential risks and advocating for global standards to ensure responsible development and deployment. He highlighted the need for proactive measures to mitigate societal impact.

Evaluating AI's safety in mental health just got harder. MentalHealthBench, a new expert-vetted benchmark, assesses helpfulness and safety across realistic conversation scenarios.

Harvey’s integration of GPT-6 Astra yields significantly improved legal document drafting quality and structure, boosting lawyer productivity by streamlining routine tasks. Initial tests show a 30% increase in draft accuracy.

GPT-6 Astra enables video platform Invideo to achieve 3x improvements in color grading and generate 50 custom effects daily, enhancing editing precision.

Ringg’s AI agents, powered by GPT-5.6, resolve up to 65% of customer calls at 90% lower cost than GPT-4.1 across multiple channels. This offers significant operational savings for businesses leveraging advanced AI.

ChatGPT Ads now available in Southeast Asia & Taiwan, extending reach to over 60 countries for participating businesses. Eligibility requirements apply.

Airbnb now offers wider internal access to GPT-6 Astra and OpenAI's newest frontier models, aiming to accelerate development workflows for engineering teams by automating tasks like bug fixing and system design.

Claude Opus 5.5 demonstrates unprecedented ability to reproduce complex physics simulations (honey coiling, simulated muscle movement) in real-time, surpassing GPT-6 Astra's capabilities; however, risks remain including potential for autonomous operation and hallucinations despite passing 16/18 quality tests.

AWS introduced an agentic AI solution for video intelligence, reducing manual review time by ~80% for a media company; users query videos via natural language, triggering services like Rekognition and Transcribe as needed, with responses under one second for cached content. The system adapts to new queries without code changes.

Gemini 3.8 text-to-speech models (Flash & Flash-Lite) now available, achieving #1 on Hume AI’s Voice Design Benchmark (71.4) and Overall Quality Index. New features include generative voice design from scratch, voice replication with consent verification, and line-by-line performance control.

DeepMind's Private AI Compute now features server-side memory, enabling persistent cross-device AI while maintaining on-device privacy standards; data remains encrypted even during cloud processing within secure enclaves. Google is also publishing verifiable software records to enhance transparency and trust.

HEMA, a Dutch retailer, slashed internal friction by deploying an AI assistant (HAL) on Amazon Bedrock AgentCore and MCP. The solution consolidated fragmented knowledge into a single source, enabling instant answers within tools like IDEs—reducing portal-hopping from afternoons to seconds. Built in two steps, HAL now supports authentication via Microsoft Entra ID, eliminating AWS credentials.