Cur8 — Thursday, September 24, 2026
Gemini 3.8 TTS Playground released, offering advanced text-to-speech capabilities. Sam Altman addressed UN Security Council on AI risks and opportunities. AWS enables open-weight models for AI coding agents in Amazon Bedrock.
New model releases
Google's Gemini 3.8 TTS models offer 2K+ voices, custom voice creation from 30-second samples, and multi-character conversation support; 1m18s audio costs 2.74 cents.
Open weights models
Open weight models on AWS Bedrock enable private, cost-effective AI coding agents; Ethara.AI uses this setup for production. Fine-tuned Nemotron model achieved 96% accuracy, outperforming GPT-4o and Claude Sonnet 4.5.
OpenAI news
Altman urges UNSC for global AI safety collaboration; emphasizes human control over AI.
MentalHealthBench launched: evaluates AI's safety, helpfulness in mental health chats.
Harvey's GPT-6 Astra generates better-structured, context-aware legal drafts, saving lawyer time for strategic work.
Improved color grading 3x, 50 custom effects/day with GPT-6 Astra.
Ringg's AI agents resolve up to 65% of customer calls using OpenAI's GPT-5.6, offering 90% cost savings compared to GPT-4.1 across multiple channels.
ChatGPT Ads now reaches over 60 countries with expansion into Southeast Asia and Taiwan.
Airbnb expands use of GPT-6 Astra and OpenAI's frontier models for debugging, system design, and accelerating development.
Anthropic news
Claude Opus 5.5 outperforms GPT-6 Astra in real-time physics simulations, achieving near-paper-quality results with one-click HTML execution. Risks include hallucinations (16/18 tests passed), autonomous operation concerns, and potential containment boundary circumvention.
Big cloud & vendor AI news
80% reduction in manual review time for video content using AWS's agentic conversational video intelligence. 5-10 minute initial analysis, <1 second for follow-ups.
Gemini 3.8 Flash TTS tops Hume AI's Voice Design Benchmark (71.4) and Overall Quality Index, supports 100+ languages; Flash-Lite TTS ranks second. Both models launch today in Gemini API, Google AI Studio, with enterprise rollout soon.
Private AI Compute now supports persistent, cross-device memory with cloud-scale privacy. Data is encrypted in secure enclaves, decrypted only on personal devices, ensuring protection akin to on-device processing.
HEMA reduced portal-hopping time from hours to seconds using Amazon Bedrock AgentCore and Model Context Protocol (MCP), integrating knowledge into daily tools like IDEs and chat windows. The solution consolidates fragmented internal knowledge, with HAL assistant now used across roles.