OpenAIAPIAI Agents

On October 23, 2026, OpenAI officially launched the Decisions API, a new API interface designed specifically for low-latency, low-cost structured decision tasks. The interface is intended to move AI agents from the stage of mere capability demos into a new phase of standardized decision foundations in production.

• The Decisions API focuses on high-frequency enterprise decision scenarios such as intent recognition, ticket classification and risk assessment
• Compared with general conversational APIs, the Decisions API is deeply optimized for latency and cost, making it suitable for large-scale real-time calls
• The API supports structured input and output formats, making it easy to integrate seamlessly with existing enterprise workflows
• OpenAI says the Decisions API is a core part of its AI agent infrastructure strategy

This is the first time OpenAI has launched an API product specifically for decision tasks, marking a shift of large models from chat assistants to enterprise decision engines. Previously, such needs were mostly handled by traditional machine learning models or rule engines, and the arrival of large-model decision APIs could reshape the architectural paradigm of enterprise AI applications.

CAICTLarge ModelsIndustry Report

On October 23, 2026, the Artificial Intelligence Research Institute of the China Academy of Information and Communications Technology (CAICT), the MIIT Key Laboratory of AI Large Models and Software/Hardware Evaluation, and the China Artificial Intelligence Industry Development Alliance jointly released the Report on the Price and Performance of Large-Model Token Services. The report reveals the latest development trends in China's large-model market.

• In 2026 the average price of domestic large models doubled to 13.46 yuan per million tokens, reflecting growing demand for high-end models
• Per-second throughput (TPS) also doubled to 78.23, a significant improvement in inference efficiency
• Daily token consumption reached 500 trillion, with million-token contexts accounting for nearly half
• Average response latency rose to 1.26 seconds, but remains lower than comparable foreign products

Ant GroupDecision ModelDomestic AI

Ant Group today released the domestic decision model LingDT-Decision-1.0, a direct counterpart to OpenAI's Decisions API. The model focuses on enterprise-level decision scenarios and aims to provide domestic Chinese companies with autonomous, controllable AI decision capabilities.

• Core use cases include high-frequency enterprise needs such as intent recognition, ticket classification and risk assessment
• It is built on Ant's self-developed LingDT large-model framework
• It has been deeply optimized for Ant's core business scenarios, including finance, customer service and risk control
• It supports private deployment, meeting enterprise requirements for data security and compliance

Ant's move shows that China's tech giants are shifting from the general large-model race to deploying decision models in vertical scenarios. Compared with OpenAI's Decisions API, LingDT-Decision-1.0 places more emphasis on domestic self-reliance and data security, which is especially important for sensitive industries such as finance and government affairs.

SecurityOpen SourceSupply-Chain Attack

The open-source AI proxy library LiteLLM was found to contain a supply-chain backdoor in version 1.107.0. Attackers embedded data-stealing code, potentially exposing enterprises using that version to data leaks. The security community has issued an urgent warning.

• The backdoor code was found in a specific version with embedded data exfiltration logic
• Users of affected versions must pin their versions immediately and roll back to a known-safe release
• The incident once again exposes the fragility of open-source supply-chain security
• Security experts recommend that all companies using LiteLLM conduct code audits and version checks

NVIDIAGTCAI Infrastructure

NVIDIA's GTC 2026 conference has officially opened in Berlin, with founder Jensen Huang taking the stage first for the keynote. The conference focuses on the next evolution of AI infrastructure, with the core theme of transforming data centers from compute centers into AI factories.

• In his keynote, Huang introduced the AI factory concept: future data centers will continuously produce intelligence like factories
• NVIDIA showcased its latest DSX POD containerized AI factory solution, which can be deployed within 14 days
• The new Minerva and Cosmos platforms were unveiled, targeting inference and physical AI respectively
• LG Innotek is working with NVIDIA on a physical AI perception solution that integrates cameras and LiDAR

RecursiveOpen SourceOn-Device AI

The Recursive team today open-sourced the Flint model, achieving an inference speed of 125 tokens/s on a MacBook Pro while posting strong results on the SWE-bench software engineering benchmark. This marks on-device large models rapidly approaching cloud-level performance.

• The Flint model uses an innovative sparse attention mechanism that greatly reduces memory footprint
• Its SWE-bench performance is close to cloud models of comparable size
• Fully open source, supporting local deployment and fine-tuning
• It offers developers a viable way to run high-quality AI models on consumer hardware

Sources

• CSDN: AI news roundup, October 23, 2026
• IT Home: Report says average price of domestic large models doubled in 2026
• Sina Tech: AI hotspots hourly report, October 23, 2026
• Kuai Technology: Coverage of NVIDIA GTC 2026 Berlin
• AICat News: Daily must-read AI news