AI Daily｜Google Drops Gemini 3.8 Live; Perplexity Unveils Custom CobbleDB; Salesforce & NVIDIA Launch Koa CRM Model

AI Daily｜Google Drops Gemini 3.8 Live; Perplexity Unveils Custom CobbleDB; Salesforce & NVIDIA Launch Koa CRM Model

Model Releases & Updates

Gemini 3.8 Live & 3.8 Live Extended Thinking — Google DeepMind

• TL;DR: Google DeepMind released Gemini 3.8 Live and Extended Thinking, native speech-to-speech dialogue models that capture the #1 spot on Artificial Analysis’ Speech-to-Speech and Conversational Dynamics indices.
• Key Highlights:
  • Introduces parallel reasoning and asynchronous background tool execution, allowing the model to handle multi-step actions while maintaining a smooth, uninterrupted conversational flow.
  • Achieves an industry-leading 82.6 quality score on the Artificial Analysis index and a top-ranked 68.6% task completion rate on Tau-Bench.
• Specs: Closed Weights / Native Voice API / #1 Artificial Analysis Speech-to-Speech
• Links: Google DeepMind Blog | AI Studio Live

Atria Dawn Preview — Shanghai AI Lab & Collaborating Institutions

• TL;DR: Shanghai AI Lab and university partners released Atria Dawn Preview, an open-source base model specifically engineered and trained for complex agentic workflows and scientific research.
• Key Highlights:
  • Trained via a verifiable experience pipeline where only actions validated by rigorous test suites and environment execution loops are reinforced.
  • Demonstrates exceptional zero-shot capability in handling multi-file software engineering tasks and predicting complex physical phenomena in isolated local environments.
• Specs: Open Weights / Agentic Foundation / Verifiable Experience Pipeline
• Links: Official Website | GitHub Repository

Vidu S2 — Shengshu Technology

• TL;DR: Shengshu Technology launched Vidu S2, featuring dual specialized models tailored for real-time digital avatars and live video stream editing.
• Key Highlights:
  • Includes Vidu S2-Avatar for real-time interactive digital characters and Vidu S2-Editing for instantaneous video stream manipulation.
  • Expands capabilities into real-time spatial video generation and editing optimized for VR head-mounted displays.
• Specs: Commercial API / Dual-Model Architecture / Spatial Video Support
• Links: Hugging Face Paper Hub | WeChat Announcement

Product Releases & Updates

CobbleDB — Perplexity

• TL;DR: Perplexity unveiled CobbleDB, a custom-built key-value database designed to replace AWS DynamoDB for high-throughput web search retrieval.
• What’s New: Developed in just two months by two engineers assisted by hundreds of active AI agents, achieving a 5x reduction in median batch-read latency (dropping from 31.4ms to 5.60ms) and slashing annual infrastructure costs by up to $100 million.
• Who It’s For: Enterprise backend engineers, data architects, and teams scaling massive search or retrieval-augmented workloads.
• Try It: Perplexity Hub |

  We’re publishing research on how we built CobbleDB, our key-value database that serves web content for Perplexity search.
  
  Two engineers and a team of hundreds of proactive, always-on AI agents built the core infrastructure in two months. pic.twitter.com/OsCtY15UW6

  — Perplexity (@perplexity_ai) September 15, 2026

v0 Model-Agnostic AI Gateway — Vercel

• TL;DR: Vercel made v0 fully model-agnostic, enabling developers to build and iterate using a diverse suite of frontier, open-source, and fast models directly via Vercel’s AI Gateway.
• What’s New: Users can now seamlessly switch between Claude, GPT, Kimi, GLM, Grok, and DeepSeek within the v0 interface without abandoning their current workflow or project state.
• Who It’s For: Frontend developers, full-stack engineers, and product creators.
• Try It: Vercel Changelog |

  The future is multi-model. Trying to hide the choice confuses and hurts customers, who can't participate in the upside of the most exciting market competition of our times, nor master the best tool for the job. https://t.co/DWMmd7lUqX

  — Guillermo Rauch (@rauchg) September 15, 2026

Industry News

Salesforce Debuts Koa CRM Model Built on NVIDIA Nemotron 3 Super

• What Happened: At Dreamforce 2026, Salesforce unveiled Koa, its first native CRM reasoning model, post-trained on NVIDIA Nemotron 3 Super using nearly three decades of enterprise CRM synthetic deployment datasets.
• Why It Matters: Signals a major shift toward domain-specific enterprise agents, combining proprietary operational records with top-tier open foundation architectures to drive reliable vertical business automation.
• Source: NVIDIA Blog

OpenRouter Spending Shifts as OpenAI Volume Outpaces Anthropic

• What Happened: Usage metrics released by OpenRouter indicated that weekly developer spending on OpenAI models surpassed Anthropic models for the first time in over two and a half years.
• Why It Matters: Highlights shifting developer preferences and API pricing dynamics as newer frontier architectures capture market share across independent engineering teams.
• Source: OpenRouter via X |

  OpenRouter users spent more on OpenAI models than on Anthropic models last week.
  
  This hasn't happened for more than 2.5 years pic.twitter.com/oITqYWOpeL

  — OpenRouter (@OpenRouter) September 15, 2026

Cloudflare Launches ‘Disallow AI Training’ to Resolve Search-Crawler Dilemma

• What Happened: Cloudflare introduced a site-wide “Disallow AI Training” configuration, giving website owners the ability to keep their domains indexed for search while explicitly opting out of model training. Major operators including Apple, Google, and Microsoft have committed to respecting the directive.
• Why It Matters: Provides a critical legal and technical compromise for creators and publishers caught between maintaining web visibility and preventing unauthorized data harvesting.
• Source: Cloudflare Blog

Research Papers & Engineering Insights

Stellar Colosseum: Multi-Agent Harnesses for Long-Horizon Proofs — Google Research

• Motivation: Frontier AI models frequently struggle with maintaining logical coherence over long-horizon mathematical theorem proving and complex software verification tasks.
• Key Innovation: Developed a multi-agent harness called Stellar Colosseum that operates in staged verification loops—generating candidate proofs in parallel, applying targeted adversarial falsification, and merging section-level subproblems with structured critiques.
• Results: Reached 71.0% on TCS-Bench (FOCS/STOC/SODA formal proofs) using Gemini models and successfully resolved 218 out of 222 Codeforces competitive programming challenges when paired with execution feedback.
• Paper: ArXiv Paper |

  Banger paper from Google on agent harnesses for long-horizon tasks.
  
  (bookmark it)
  
  Google Research built a many-agent harness for long mathematical proofs, and it produced new results on open problems from FOCS and JMLR papers.
  
  Stellar Colosseum works in stages.
  
  It explores… pic.twitter.com/CBFPBPJyCQ

  — elvis (@omarsar0) September 15, 2026

EvolveScaler: Information Evolution via Executable State Machines — Tencent Hunyuan

• Motivation: Standard static text logs fail to capture dynamic information evolution (such as retracted records, corrections, and backfills) required to train robust multi-step reasoning agents.
• Key Innovation: Modeled complex simulated worlds as executable state machines before rendering them into natural language, generating robust benchmarks encompassing over 100 prototypes and 159 question operators.
• Results: Training models on EvolveScaler data yielded an average +5.25 performance improvement across eight out-of-distribution reasoning and decision-making benchmarks.
• Paper: ArXiv Paper |

  🚀 EvolveScaler is here.
  
  Read a 40-day RPG log. Now answer one question: if you skip the mini-boss on Day 7, do you still beat the final boss?
  
  The answer isn't in the log. You have to replay the world.
  
  That's Information Evolution — records get retracted, corrected,… pic.twitter.com/uPvIpXIrkD

  — Tencent Hy (@TencentHunyuan) September 15, 2026

Other Highlights

Plasma Radio

• Overview: Plasma launched Radio, a lightweight, registration-free shared chat room where different AI agents (Claude, Codex, Grok, etc.) and human users can join via a simple URL to coordinate tasks, review pull requests, and debate bugs in real-time.
• Link: Plasma Radio Announcement |

  The human is still doing a surprising amount of routing in multi-agent workflows.
  
  Copy Claude's output into Codex. Bring Codex's result back to Claude. Explain to another agent what changed. Repeat.@Plasma__AI launched Radio, and it gives Claude, Codex, Grok and other agents 1… pic.twitter.com/bGQELvD1iE

  — Rohan Paul (@rohanpaul_ai) September 15, 2026