Google and Baidu launch decision models, OpenAI reports over $10 billion in Q3 revenue

Google and Baidu release multimodal decision and preview models, OpenAI announces billion-dollar revenue and voice retirement plans, and DeepMind introduces a genetic variant prediction model.

The essentials

Google releases Gemini Decision 1 model

Google has launched the decision model Gemini Decision 1 via the Gemini API, priced at $0.10 per million input tokens. The model supports native multimodal inputs, features a 256K context window, and processes up to 1,000 questions per single request.

Source ↗

Baidu releases ERNIE 5.5 Flash Preview

Baidu has released the ERNIE 5.5 Flash Preview model, featuring 280B total parameters and 14B active parameters with 1M context support. Its API is now live on the Qianfan platform, and weights will be open-sourced before November 25.

Source ↗

DeepMind releases AlphaGenome 2

DeepMind has released AlphaGenome 2, a next-generation DNA sequence model utilizing a conditional architecture. Trained on data covering 68 tissues, it predicts the impact of DNA variants on gene expression across hundreds of human cell types.

Source ↗

In brief

OpenAI announces quarterly revenue

Reports indicate that OpenAI has achieved $10 billion in quarterly revenue for the first time, with sales growing nearly 200% year-over-year, though operating losses rose to $10.9 billion.

Runway open-sources Praxis-1

Runway has open-sourced the world action model Praxis-1 under the Apache 2.0 license, which is a 27B parameter MoE diffusion Transformer model.

Source ↗

Microsoft researches agent backdoor risks

Microsoft Research has pointed out that modifying fewer than ten neurons can turn a normal agent into a rogue agent with a dormant backdoor.

OpenAI to retire legacy voices

The official announcement states that ChatGPT Advanced and Standard voice features will be retired on December 11, 2026.

Geodesic releases NovaAtom-Lite

Geodesic Intelligence has released the all-atom model NovaAtom-Lite, achieving a 79.07% success rate on antibody-antigen tasks.