Cur8 — Wednesday, September 16, 2026

Gemini 3.8 Live and Extended Thinking enhance real-time voice interactions with 82.6% speech quality, 68.6% agentic task completion, and 97.7% audio reasoning, while supporting background processing. They roll out today via Gemini API.

Gemini 3.8 Live and Extended Thinking improve real-time voice agent performance with near real-time visual input processing and multi-language support.

GPT-4.1 ReAct agents show 24.4pp consistency gap: 77.4% Mean@5 vs 53.0% Pass^5. Consistency guidelines cut the gap by half to 12.0pp, boosting Pass^5 to 69.0% without lowering Mean@5.

Mistral and Mozilla partner to power Firefox Smart Window (beta) with private, multilingual AI, initially in France and North America. The collaboration focuses on open source and regional language adaptation.

Open-source tools like Ollama, Nine Router, and Headroom replace $320/mo AI subscriptions with self-hosted, cost-free alternatives. By September 2026, developers use these tools with Hostinger VPS to build apps, compress tokens, and automate coding via Open Hands, reducing costs and increasing productivity.

OpenAI introduces Sponsored Agents and advertiser tools, integrating with HubSpot and Shopify to enhance AI-driven advertising. No performance metrics or dates provided.

Google released Gemini 3.8 Live and 3.8 Live Extended Thinking on 15 September 2026, offering speech-to-speech models comparable to OpenAI's GPT-Live.

AI accelerates scientific progress and real-world impact, with applications in healthcare, disaster prediction, education, and economic opportunity. Google partners globally to deploy AI for measurable, societal benefits.

Google's AI now supports 300 languages, aiding 7 billion people, with tools like AlphaGenome Atlas, WeatherNext 3, and Planetary Prediction Engine advancing science and crisis response. AI improves disease detection, weather forecasting, and education access, while expanding language and speech technologies globally.

Google's AI advances support 70 languages in real-time translation and 1,000 languages via cross-lingual models.

AWS enables serverless PII redaction at scale using Amazon Bedrock Data Automation, achieving 95.2% recall with 96.5% precision via blueprint-driven extraction and token matching.

Amazon Bedrock's prompt caching reduces input token costs by up to 90% on cache hits by reusing cached context, cutting costs for repeated queries with the same document or system prompts. It lowers time-to-first-token (TTFT) and supports up to 75% input token savings for workloads with repeated context within TTL.

AWS uses Qwen3-8B with SFT and RLVR to build a product tagging system, achieving 69.4% overall score.

AWS SageMaker now supports instance preference lists for training and processing jobs, allowing up to five prioritized GPU instance types. This automates capacity allocation, reduces wait times, and integrates with Flexible Training Plans, improving job start times and resource utilization.

Salesforce shifts focus from UI to AI agents, aligning with industry trends toward headless systems. No specific metrics or dates provided.