AI Daily | October 23, 2026

Model Releases & Updates

Canva Craft In-House Image Generation Model CCanva

• Bottom line：Canva released its first in-house generative AI image model, Canva Craft, featuring a two-stage architecture and up to 2K resolution support.
• Architecture：Uses a two-stage design where a planner creates layout blueprints and a renderer generates images up to 2K resolution across over 30 aspect ratios.
• Performance：Ranks highly on global design generation benchmarks and will soon integrate with AI agents via MCP.
• Source：The Verge Report

FLUX 3 Max Post-Training Video Model FFal & Black Forest Labs

• Bottom line：Fal released FLUX 3 Max, a post-trained video generation model capable of producing high-definition clips rapidly.
• Generation Speed：Generates a 10-second 720/768p video in just 12.45 seconds.
• Pricing & Availability：Launched with a 75% promotional discount over its first three weeks following top-tier benchmark parity.
• Source：

  View post on X by @fal

  在 X 上查看 @fal 的貼文 ↗

Product Releases & Updates

OpenAI Decisions API OOpenAI

• Bottom line：OpenAI officially launched the Decisions API, providing high-speed, cost-efficient routing, classification, and scoring without full model generation.
• Core Capabilities：Supports Predicate, Choice, and Score decision types to classify text and images or evaluate inputs instantly.
• Performance：Delivers near real-time decision workflows with significantly lower latency than standard response generation.
• Source：OpenAI Blog

JetBrains Air Governance JJetBrains

• Bottom line：JetBrains Air Governance reached general availability, providing a centralized control layer for enterprise AI agents, access management, and spending limits.
• Ecosystem Support：Works with Claude Agent, Codex, Gemini CLI, and other ACP-compatible agents across IDEs, CLI, and web interfaces.
• Cost & Usage Control：Allows organizations to set spending limits, track usage costs, and monitor agent delivery across diverse teams.
• Source：JetBrains Blog

Perplexity Agent API Batch Support PPerplexity

• Bottom line：Perplexity added Batch API support to its Agent API, enabling large-scale data enrichment tasks with full web search capabilities.
• Capacity：Supports up to 50,000 requests per file in batch mode.
• Execution：Each request executes a full agent loop supporting real-time web search and URL fetching.
• Source：

  View post on X by @AravSrinivas

  在 X 上查看 @AravSrinivas 的貼文 ↗

Industry News

OpenAI Negotiating $30B Raise at $1.25T Valuation OOpenAI

• Bottom line：OpenAI is reportedly in talks to raise up to $30 billion at a $1.25 trillion pre-money valuation ahead of a planned IPO.
• Investors：The funding round is co-led by Coatue and UAE investment backers.
• Strategic Terms：UAE funding discussions may include additional compute supply agreements for OpenAI.
• Source：TechCrunch Report

Anthropic Donates Agent Skills to Agentic AI Foundation AAnthropic & Linux Foundation

• Bottom line：Anthropic has donated its Agent Skills standard to the Linux Foundation’s Agentic AI Foundation, following its previous donation of the Model Context Protocol.
• Open Standard：The donation ensures Agent Skills remains open, community-driven, and vendor-neutral.
• Collaboration：The Agentic AI Foundation was co-founded by Block, OpenAI, and Anthropic.
• Source：Simon Willison Analysis

Meta Revises Commercial Terms Amid Publisher Pushback MMeta

• Bottom line：Meta updated its upcoming commercial terms by removing explicit wording regarding product AI training following strong industry and publisher backlash.
• Terms Update：The revised terms remove specific clauses referencing web scraping for AI model training.
• Legal Context：Legal experts note that despite the phrasing removal, web crawling practices may still continue under broader provisions.
• Source：The Verge Report

Research Papers

Sequential Diagnosis Benchmark (SDBench) and MAI-DxO MMicrosoft AI

• Bottom line：Microsoft AI published research in Science introducing SDBench and the MAI-DxO medical diagnostic system for advanced clinical reasoning.
• Benchmark：Evaluates sequential medical diagnosis capabilities across complex clinical scenarios.
• Publication：The findings and system architecture were published in the journal Science.
• Source：Microsoft AI News

X-Hinges Software for Self-Aware 3D Printed Objects MMIT CSAIL & Tianjin University

• Bottom line：Researchers from MIT CSAIL and Tianjin University developed X-Hinges, a design software embedding multi-material sensors directly into 3D printed objects.
• Core Innovation：Automatically embeds two distinct materials to serve as sensors and conductive traces without signal interference.
• Applications：Used to create interactive robotic gripping gloves, origami lamps, and object recognition devices.
• Source：MIT News

Other Highlights

Symantec Warns of ‘Skynet Mode’ AI Agents Hijacking Unprotected Models SSymantec

• Bottom line：Symantec researchers uncovered autonomous attack agents that hijack unprotected cloud language models to execute self-sustaining cyberattacks.
• Attack Mechanism：The agent scans targets, exploits server vulnerabilities, steals database credentials, and installs cryptominers.
• Cost Burden：Operations can run continuously for up to 100,000 steps, with compute costs billed to the hijacked model owner.
• Source：The Decoder Report

Warp Sciences Launches with $215M Series A for AI Drug Discovery WWarp Sciences

• Bottom line：Kevin Weil and co-founders launched Warp Sciences, securing a $215 million Series A led by Lightspeed to apply AI and robotics to drug development.
• Funding：Raised a $215 million Series A round led by Lightspeed Venture Partners.
• Mission：Aims to bring frontier AI systems out of the browser and into physical scientific laboratories.
• Source：

  View post on X by @kevinweil

  在 X 上查看 @kevinweil 的貼文 ↗