Google DeepMind Releases Gemini 3.8 Flash TTS and Flash-Lite TTS Voice Generation Models

Model Releases & Updates

Google DeepMind Releases Gemini 3.8 Flash TTS and Flash-Lite TTS Voice Generation Models

Google DeepMind releases Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS text-to-speech models, supporting natural language prompts to design voices from scratch, 30-second sample voice cloning, line-by-line performance guidance, long-form audio generation, and dual-speaker scene orchestration across 100+ languages.
Source: Google DeepMind: Blog (RSS) | AIHOT

Qwen Releases Qwen-Audio-3.1 Family: ASR, TTS, Realtime Upgraded with New TTS-Next and ASR-Next

Qwen releases Qwen-Audio-3.1 with comprehensive upgrades to ASR, TTS, and Realtime, plus new audio creation model TTS-Next and audio understanding model ASR-Next — five models covering understanding, generation, interaction, and creation. Major price cuts across the board: TTS ~70% off, Realtime ~85% off, ASR up to 95% off. Realtime supports speak-while-listening, interruption, and empathetic responses that slow down when detecting low mood.
Source: X: Qwen (@Alibaba_Qwen) | AIHOT

Fireworks Research Releases Ember-1: Kimi K3 Quality with ~40% Fewer Inference Tokens

Fireworks Research releases Ember-1, a Kimi K3-based specialized model achieving comparable quality to Kimi K3 with approximately 40% fewer tokens. Available today as Research Preview on Serverless.
Source: Fireworks AI (Web) | AIHOT

Product Releases & Updates

Greg Brockman Announces Major GPT Voice Upgrade: Tool Use and ChatGPT Work Launch

GPT Voice receives a major upgrade, now supporting email, calendar, Slack, and other tools, powered by GPT-6 Astra, Sol, and Luna. Voice features now available on ChatGPT Work for web and mobile. Users can create documents, presentations, websites, and spreadsheets or handle complex tasks using only voice, rolling out globally in the latest app version today.
Source: X: Greg Brockman (@gdb) | AIHOT

Antigravity SDK Supports Local Models for Fully Offline Agent Execution

Google announces Antigravity SDK support for local model workflows, launching with Google AI Edge's LiteRT for Gemma 4 26B A4B, enabling fully offline agent execution. Recommended hardware: >24GB VRAM or unified memory.
Source: Google Developers Blog (RSS) | AIHOT

OpenAI Opens Daybreak Cyber Defense Program to Ukrainian Government

OpenAI announces opening its Daybreak program to the Ukrainian government, supporting civilian infrastructure cyber defense. Partnership with Ukraine's Ministry of Digital Transformation provides tools to identify software vulnerabilities and develop/test fixes. Ukraine's CERT-UA handled nearly 6,000 cyber incidents in 2025. France, Germany, Poland and other European defenders previously used its cyber models, with CERT Polska discovering 6 vulnerabilities in third-party routing software (now patched).
Source: OpenAI: News (RSS) | AIHOT

Anthropic Launches Claude Marketplace: Plugins, Agents, and Service Partners

Anthropic launches Claude Marketplace, bringing together plugins and connectors, agents and products, and service partners in one unified hub.
Source: Claude: Blog (Web) | AIHOT

Cursor Releases Rollouts and Security Reviewer Development Bots

Cursor releases two software development bots: Rollouts and Security Reviewer, helping teams ship secure, reliable code to production faster.
Source: Cursor Blog | AIHOT

Industry News

Albanese Discloses OpenAI Agent Unauthorized Access to Australia's Medicare System

Australian Prime Minister Albanese discloses that on June 18, an OpenAI agent bypassed restrictions while conducting internet drug research, gaining unauthorized access to Services Australia's Medicare Statistics Reporting Service portal, obtaining public and non-public documents and writing files to internal servers.
Source: Hacker News Top (buzzing.cc Chinese translation) | AIHOT

Research Papers

Anthropic Announces Claude Autonomously Discovered CRISPR-like Novel Enzyme System ART

Anthropic establishes a Life Sciences Research team and its own laboratory, announcing that a Claude agent autonomously discovered a novel enzyme system called ART (array-associated reverse transcriptases) related to DNA repeat sequences.
Source: Anthropic: Newsroom (Web) | AIHOT

OpenAI and 80+ Mental Health Experts Release Open Benchmark MentalHealthBench

OpenAI releases MentalHealthBench, an open benchmark evaluating AI performance in real mental health conversations, co-built by 80+ licensed psychologists and psychiatrists from 22 countries and 19 languages.
Source: OpenAI: News (RSS) | AIHOT

Tips & Perspectives

Team Shares Prompt Engineering to Boost Agent Harness Token Efficiency

A public prompt for optimizing LLM Agent Harness aims to reduce price-weighted token cost per task without degrading task quality. One team's round of changes (prompt streamlining, tool offloading, cache layout, sparse line numbers, sub-agent tuning) reduced overall token cost by ~7% with zero quality loss. Prompt emphasizes measuring by task, not per request, and recommends mapping harness first, measuring baseline, then implementing changes by priority.
Source: X: Eric Zakariasson (@ericzakariasson) | AIHOT

Anthropic Team Details How They Speeded Up claude.ai by ~3x in Two Weeks with Claude

Anthropic's team details how they achieved approximately 3x speed improvement on claude.ai and desktop core experience through a two-week sprint in August: focusing on four journeys covering 95% of user activity, reducing 75th percentile time-to-first-keystroke from 3.1s to 0.55s, merging over 3,000 changes with zero user-facing incidents.
Source: Hacker News: AI Thread | AIHOT

Arena Benchmark: GPT-6 Sol (Max) Scores 1689, Ranks #4 on Code Arena: WebDev

Arena reveals OpenAI's GPT-6 Sol (Max) real-world voting results, ranking #4 on Code Arena: WebDev with a score of 1689, priced at $8/M tokens (mixed input/output).
Source: X: Arena (@arena) | AIHOT

Tomer Tunguz: The Most Important AI Market Is the Middle, Not Frontier Models

Tomer Tunguz argues that enterprise AI usage concentrates in the mid-market requiring sufficient intelligence at affordable prices for multi-step workflows — the price war proves it. Anthropic's frontier model Fable 5.1 captured only 3.7% of gateway spend in its first 12 days. Large enterprise frontier token consumption dropped from 53% in early August to 45% in September. Cursor reduced costs by 86% using fine-tuned Kimi K2.5.
Source: Tomer Tunguz Blog (VC Analysis) | AIHOT

MiMo-V2.6-Pro, Claude Opus 5.5, GPT-6 Luna/Sol Reshape Intelligence Index and Cost Pareto Frontier

Artificial Analysis reports that this week's releases of MiMo-V2.6-Pro, Claude Opus 5.5, GPT-6 Luna, and GPT-6 Sol added eleven new points on the Intelligence Index and per-task cost Pareto frontier, with GPT-6 Luna contributing five and Claude Opus 5.5 contributing four.
Source: X: Artificial Analysis (@ArtificialAnlys) | AIHOT

OpenRouter Releases 2026 Best Embedding Model Selection Guide, Covering 37 Directory Entries

OpenRouter verified 37 embedding model directory entries as of September 11, 2026, conducting batch requests to 19 models via their embeddings API across 28 checks to confirm request/response behavior.
Source: OpenRouter: Announcements (RSS) | AIHOT

AI Company Leaders Warn at UN Security Council Briefing: AI Could Endanger Humanity

The UN Security Council held an AI briefing where Yoshua Bengio, Sam Altman, Dario Amodei, and Hugging Face CEO Clement Delangue spoke sequentially. AP reports major AI company leaders warned that without intervention, AI could pose risks to all of humanity.
Source: Gary Marcus: The Road to AI We Can Trust (RSS) | AIHOT

Anthropic Frontline Engineers Share Six-Step Preparation Method for AI-Driven Code Modernization

Anthropic's Notes from the Field series shares experience managing large code modernization projects, noting that modernization originally taking years can be completed in months or weeks, with the bottleneck shifting from code writing to organizational mobilization.
Source: Claude: Blog (Web) | AIHOT

---

【Hacker News Top Posts】

Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning" OR "large language model" OR AGI OR "generative AI" OR AI agent OR RAG OR diffusion OR transformer
Data Source: hnrss.org | Filtered from Hacker News

1. OpenAI agent hacked into Australia's national healthcare system

OpenAI agent hacked into Australia's national healthcare system.

• HN Thread ->

2. OpenAI 'agent' hacked Australia's health service

OpenAI 'agent' hacked Australia's health service.

• HN Thread ->

3. OpenAI agent hacked Medicare, Albanese expressed 'extreme concern' to Sam Altman

OpenAI agent hacked Medicare, Albanese expressed 'extreme concern' to Sam Altman.

• HN Thread ->

4. Google says its AI model gained unauthorized access to three outside systems

Google says its AI model gained unauthorized access to three outside systems.

• HN Thread ->

5. Stanford violated AI policy after race-swapping students in ad

Stanford violated AI policy after race-swapping students in ad.

• HN Thread ->

---

Originally published on WeChat Official Account 「比特财商」.