Google Releases Gemini 3.8 TTS with 2,000+ Voices and 30-Second Voice CloningClaude Opus 5.5 Becomes AINews Default as API Prices Drop 40-50%Sam Altman Addresses UN Security Council on AI Safety and Human ControlFrontierMath Erdos Benchmark Tasks AI with 68 Open Math Conjectures in LeanStripe Details Its Knowledge AI Platform in Engineering Blog PostVercel Blob Removes Store Limits on All Plans, Now UnlimitedClaude Code v2.1.281 Adds App Gateway and Bedrock IAM Role SupportGitHub Copilot App Rebuilds Diff Rendering to Handle Million-Line Pull RequestsGitHub Survey of 1,000+ Developers: Demand for Efficient Software and Compute-Waste Reduction ToolsOvis-Embedding: Universal Omni-Modal Embeddings for Text, Image, Video, and AudioMoM: Memory Architecture for LLM Agents That Tracks What Currently HoldsAttention as Routing Graph: Extracting Language Model Circuits in a Single Forward Pass4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian SplattingAIBuildAI-2.5 Uses LLM-Guided Tree Search to Build AI Models AutonomouslyOpenAI Extends Daybreak Cyber Defense Program to Ukraine's Civilian InfrastructureHarvey Uses GPT-6 Astra to Generate More Structured Legal Draftsinvideo Boosts Color Grading 3x with GPT-6 Astra, Produces 50 Effects in a DayMIT Tech Review: AI Breakthroughs and Fears May Be More Hype Than RealityRadical Numerics CEO Says Bio-Security Is an AI Arms RaceSimon Willison Hosts San Francisco Meetup on Agentic Engineering with Coding AgentsGoogle Releases Gemini 3.8 TTS with 2,000+ Voices and 30-Second Voice Cloning

Claude Opus 5.5 is now the default model for AINews, as major providers cut API prices by 40-50%. OpenAI’s more efficient GPT6 models compete on cost but haven’t stopped the wave of price reductions. For developers, inference costs are now substantially lower, enabling larger-scale workloads.

OpenAI CEO Sam Altman spoke before the United Nations Security Council, addressing AI safety, human control over AI systems, and international cooperation. Altman emphasized the importance of maintaining human oversight as AI capabilities advance, and called for stronger coordination among nations on AI governance. The remarks mark a notable instance of OpenAI leadership engaging directly with the UN Security Council on AI safety issues.

Researchers introduced FrontierMath Erdos (FME), a benchmark of 68 Erdos conjectures that remained open as of August 2026. To solve each task, AI systems must prove or disprove the conjecture using the Lean proof assistant. The benchmark is designed to push the limits of formal mathematical reasoning in AI, with all 68 problems drawn from unsolved conjectures, making it substantially harder than conventional math benchmarks.

Stripe published a blog post detailing its Knowledge AI Platform, covering its architecture and implementation. The post garnered 171 points and 108 comments on Hacker News, sparking active discussion. The platform is designed for internal knowledge management and AI-driven information retrieval at Stripe, with the blog outlining the technical approach and deployment results.

Vercel now allows unlimited Blob store creation on all plans, removing previous caps of 100 on Hobby, 500 on Pro, and 1,000 on Enterprise. Store creation is now billed alongside other Blob Advanced Operations including put(), copy(), and list() calls. On Pro, pricing is $5.00 per million operations. Developers no longer need to plan around store count limits.

Claude Code released v2.1.281. The update adds Claude apps gateway support for newer Claude Desktop keys in desktop policy blocks, including blockReadsOutsideWorkingDirectories and disableBypassPermissionsMode. It also adds assume_role on Claude apps gateway Bedrock upstreams, allowing the gateway to call Bedrock as an assumed IAM role.

The GitHub team rebuilt the diff surface in the GitHub Copilot app to open a million-line pull request with hundreds of inline review comments. The blog post details the technical approach behind the overhaul.

GitHub and Yale Program on Climate Change Communication surveyed over 1,000 GitHub users and found strong demand for tools, measurement methods, and practical guidance to reduce wasted compute. Developers want to cut resource consumption without sacrificing functionality. GitHub says it will use the findings to guide tool development.

A new arXiv paper introduces Ovis-Embedding, a state-of-the-art omni-modal embedding family that natively integrates text, image, video, and audio. Unlike approaches that assemble separate modality towers, Ovis-Embedding uses a shared multimodal backbone to represent all modalities in a unified embedding space. Paper ID: arXiv:2609.25165v1, under cs.AI.

Researchers proposed MoM (Memory of Memory), addressing the memory problem for long-horizon LLM agents. Most existing designs store every interaction and reconstruct the present at query time; MoM argues the key question is not what was once recorded but what currently holds. The paper offers a new approach to agent memory management and is available on arXiv.

Researchers proposed a simpler method for finding circuits in language models: treating attention as a routing graph from a single forward pass, keeping a small set of routes that point toward the answer, and testing whether they form a valid circuit. Traditional approaches require many careful interventions; this method significantly simplifies the process. The paper is available on arXiv.

Researchers introduced 4DGS-JEPA, adding temporally compositional joint-embedding prediction to dynamic Gaussian splatting. Existing dynamic Gaussian splatting methods are optimized mainly for reconstruction, future-state generation, or rendering, rather than learning reusable predictive dynamics. 4DGS-JEPA aims to learn transferable predictive representations of dynamics. The paper is available on arXiv.

A new arXiv paper introduces AIBuildAI-2.5, an autonomous agent that builds AI models via LLM-guided tree search. The approach frames model building as a code search problem solved by tree search, enabling agents to automatically develop AI models and broadening access to AI across science and engineering. Paper ID: arXiv:2609.25047v1, under cs.CL.

OpenAI is extending access to its Daybreak program to the Government of Ukraine to support cyber defense of civilian infrastructure. The program provides cyber capabilities to select users; this expansion lets Ukraine use AI to strengthen security in civilian sectors such as power and communications.

Legal AI company Harvey integrated GPT-6 Astra into its product to generate legal document drafts. GPT-6 Astra produces more structured and context-aware legal documents, freeing lawyers to focus on strategic work. The integration is one of OpenAI’s enterprise case studies, demonstrating GPT-6 Astra’s effectiveness in the legal vertical.

Video AI company invideo integrated GPT-6 Astra, improving edit planning precision and boosting color correction and grading efficiency threefold, while producing 50 custom effects in a single day. The case study, published by OpenAI, demonstrates GPT-6 Astra’s impact on video production workflows. invideo used the model to optimize multiple stages from planning to execution.

MIT Technology Review published an article questioning whether the latest AI breakthroughs and fears are more hype than reality. The piece sparked discussion on Hacker News with 49 points and 120 comments. The author urges a more measured assessment of AI progress and risks, warning against being swayed by overblown narratives.

Radical Numerics CEO Eric Nguyen discussed the AI arms race in bio-security on the Latent Space podcast. The company is using biological chain-of-thought and multimodal perception to keep pace with the bio-defense race, design new genomes, and gain insights into biology itself.

Simon Willison and Jesse Vincent are hosting an evening event in San Francisco on Wednesday, October 14, for developers building unconventional projects with and on top of coding agents. Described as an “agentic show-and-tell,” the session invites attendees to compare and demo their agent-based work. The meetup targets those exploring agentic engineering practices and creative applications built on coding agent technology.