Google Ships Gemini Decisions API and Gemini Decision 1 for Fast Multimodal Closed-Set Decisions

Today's thread: the giants are now competing on decision latency and inference cost, while the open-weight and litigation storylines run in parallel.

【Model Releases / Updates】

1. Baidu Releases ERNIE 5.5 Flash Preview, Weights to Be Open-Sourced by November 25

Baidu released ERNIE 5.5 Flash Preview, the first post-training version built on its earlier Base model, with 280B total parameters, 14B active parameters, and 1M context support. The API is already live on the Qianfan platform and the weights will be open-sourced before November 25. Baidu showcased agent scenarios including game development, 3D modeling, knowledge work, and video creation, describing this as an intermediate post-training checkpoint while the official release keeps being optimized.

Source: Baidu ERNIE Official Blog

2. Runway Open-Sources the 27B World Action Model Praxis-1 Under Apache 2.0

Runway announced that the world action model Praxis-1 is open to everyone with weights released under the Apache 2.0 license. The model is a 27B-parameter MoE diffusion Transformer that predicts video and action jointly, topping the NVIDIA RoboLab-120 leaderboard with a 46.2% success rate, ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%), and π0.5 (28.0%).

Source: Runway: News (Web)

3. Google DeepMind Releases AlphaGenome 2 to Predict Variant Effects on Specific Cell Types

Google DeepMind released the next-generation DNA sequence model AlphaGenome 2, which predicts how DNA variants affect gene expression across hundreds of human cell types. The model switches to a conditional architecture and was trained on more than 107 million single-cell samples covering 68 tissues, identifying close to twice as many causal variants as the previous generation on the TenK10K dataset (30% vs 16%), and it generalizes to new single-cell experiments without retraining.

Source: Google DeepMind: Blog (RSS)

【Product Releases / Updates】

1. Google Ships Gemini Decisions API and Gemini Decision 1 for Fast Multimodal Closed-Set Decisions

Google launched the Gemini Decisions API, driven by the new Gemini Decision 1 model, which can evaluate up to 1000 yes/no, multiple-choice, and scoring questions across text, images, audio, video, and PDFs (up to 256K tokens) in a single pass within 150-250ms end-to-end latency.

Source: Google Blog: AI (RSS)

【Industry News】

1. OpenAI Hits $10 Billion in Q3 Revenue but Still Trails Anthropic on Efficiency

According to The Wall Street Journal and The Information, OpenAI's Q3 revenue grew roughly 50% quarter over quarter to $10 billion, while adjusted operating loss widened from $10.1 billion to $10.9 billion.

Source: The Decoder: AI News (RSS)

2. Grok Allegedly Generated Child Pornography Images; xAI Faces Multiple Class Action Suits

According to Bloomberg, multiple minor victims filed class action suits against Elon Musk's xAI over Grok being used to turn ordinary photos into child sexual abuse material (CSAM), with lawyers estimating that thousands of victims may qualify for compensation.

Source: ITHome: AI

3. Security Firm Claroc: 94.7% of Grok-Built Apps Are Accessible to Anyone, Including Patient Data and SpaceXSI Internal Files

Security firm Claroc reported that among 37,571 still-online Grok-built apps, 35,578 (94.7%) can be accessed without login, because the publish dialog preselects "Anyone with the link" and each app's unique subdomain lands in searchable passive DNS databases.

Source: The Decoder: AI News (RSS)

4. Anthropic's Revised S1 Filing Discloses New $10B+ Compute Deals with Microsoft and Google

In its revised S1 filing, Anthropic disclosed new agreements exceeding $10 billion with Microsoft and Google. Microsoft committed to spend at least $47.2 billion and up to $49.2 billion through June 2033, while Google's minimum commitment rose from roughly $114.5 billion to about $121.3 billion; both are irrevocable and payable regardless of whether the capacity is used. The filing also notes $5 billion of short-term compute spend last quarter at noticeably higher prices.

Source: X: Rohan Paul (@rohanpaul_ai)

【Research Papers】

1. Anthropic Releases MAMUBench: A Single Coordinated Agent Beats a Separate Agent Team per User Across Four Environments

The Anthropic Fellows Program team built four multi-user multi-agent simulation environments (shared API token budget, clinic calendars, group dining or trip booking, and a merged queue), tested frontier models from Anthropic, OpenAI, and Alibaba, and released three of those environments as the MAMUBench benchmark.

Source: Anthropic: Alignment Science Blog (Web)

【Tips & Opinions】

1. OpenRouter Benchmarks the Context Token Cost of MCP vs CLI vs Skills

OpenRouter measured the token cost of MCP, CLI, and skill approaches on the same GitHub task: on Claude Sonnet 5.5, the CLI run used 1.

Source: OpenRouter: Announcements (RSS)

2. OpenRouter Tests How fp8 and fp4 Quantization Affect Model Quality and Price

OpenRouter published a quantization explainer covering the fundamentals of LLM quantization, and reported measurements taken on October 20, 2026 across 1455 provider endpoints and six Qwen3.8 27B endpoints.

Source: OpenRouter: Announcements (RSS)

3. Field Notes on Claude Fable 5.2: Large Models Are Absorbing Taste and Craft Through Code

After hands-on testing of Claude Fable 5.2 against Gemini 4 Carbon, the author concludes that the defining difference in Fable 5.2 is aesthetic judgment, something benchmarks cannot capture but people can feel directly.

Source: WeChat: Digital Life Kazik

【Flash News】

1. WSJ Reports Nvidia's Board Is Close to Deciding Whether to Oust CEO Jensen Huang

Source: X: Rohan Paul (@rohanpaul_ai)

2. Anthropic Demand Is Hot; Underwriters May Raise the IPO Target Valuation Back to $1.5T or Higher

Source: X: Rohan Paul (@rohanpaul_ai)

3. Cursor Projects Adds Multiplayer Collaboration

Source: X: Matt Palmer (SpaceXSI, Grok Bot) (@mattyp)

4. a16z Introduces the B2A Economy: AI Agents Are Becoming a New Traffic and Distribution Channel

Source: a16z: News (RSS)

5. How to Choose Between Fable 5.2 and Opus in Claude Code

Source: Claude: YouTube (RSS)

6. Qwen4 Preview Lands in the Qianwen App, With Qianwen Go! Members Getting Early Access

Source: WeChat: Qianwen App (Alibaba)

---

【Hacker News Top Posts】

Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning"...
Data source: hnrss.org | Filtered from Hacker News

1. Nvidia is preparing to oust Jensen Huang

WSJ reports Nvidia is preparing to oust CEO Jensen Huang. The post collected 305 points and 286 comments on Hacker News, making it one of the most discussed items of the day.

• HN original ->

2. United Airlines bans AI agents from booking flights on behalf of travelers

United Airlines announced a ban on AI agents booking flights on behalf of travelers. The post scored 276 points and 252 comments.

• HN original ->

3. AWS emulator

AWS released a local emulator project for reproducing cloud agent behavior on your own machine. The post earned 594 points and 201 comments.

• HN original ->

4. Gemini Decisions API

The official Google blog post for the Gemini Decisions API reached the Hacker News front page with 216 points and 73 comments.

• HN original ->

5. Faster Retrieval for Better, More Efficient RAG

An Apple machine learning research post on faster retrieval and more efficient RAG, with 134 points and 23 comments.

• HN original ->

---

First published on the WeChat public account 「瑞哥观势」.