Give AI useful work, with meaningful controls

Decisions API goes GA for frequent, bounded decisions

2026-10-22 · New

OpenAI made its GPT-6 Luna-based Decisions API generally available. It supports probability, choice and score outputs, up to 400 questions per request, native GPT-Live integration, prompt caching and US/EU inference residency.

It targets classification, routing and next-action selection. OpenAI’s up-to-10x speed claim compares against Luna through the Responses API, not every workload. Evaluate accuracy, latency and the cost of wrong decisions on your own tasks.

OpenAI: Decisions API general availability

Claude Code 2.1.309 repairs sandbox and permission checks

2026-10-23 · New

The stable release fixes shell startup files running outside the sandbox, macOS symlink handling for denyWrite rules, and sandboxed commands or auto-accepted edits modifying a specified MCP configuration file.

These fixes affect whether automated execution respects its constraints. Teams using these modes should assess the update and retest deny rules. Release notes establish delivery of fixes, not the elimination of every possible bypass.

Claude Code v2.1.309

Thousands of agents analyze brain-cell data to generate testable leads

2026-10-21 · Recent

OpenAI, PerturbAI and Scripps Research report using thousands of GPT-Rosalind-powered agents to analyze existing mouse-brain experimental data and connect it with literature in 32 minutes, producing an explorable reasoning atlas.

The value is broader interpretation of experimental data and better prioritization of follow-up experiments. Timing and performance are team-reported, and proposed mechanisms still need experimental validation; this is not a human treatment discovered in 32 minutes.

OpenAI: Distributed biological reasoning

Claude Code · Anthropic

v2.1.309 · 2026-10-23 · stable

Stable: Fixes shell startup, symlink write restrictions and MCP configuration protection. See AI Companies for implications.

Claude Code v2.1.309

Checked 2026-10-23T03:55:32.096644+00:00

Qwen Code · Qwen

v0.26.0-nightly.20261022.bf66a26c97 · 2026-10-23 · nightly

Nightly prerelease: Adds daemon-token persistence and durable manual context compression for managed agents. It does not replace stable 0.26.0; production use needs separate validation.

Qwen Code nightly 20261022

Checked 2026-10-23T03:55:32.096644+00:00

Decisions API · OpenAI

GA / GPT-6 Luna · 2026-10-22 · stable

Generally available: Up to 400 questions per request, GPT-Live integration, prompt caching and US/EU inference residency.

OpenAI: Decisions API general availability

Checked 2026-10-23T03:55:32.096644+00:00

Beyond memory bandwidth, ask how far the data must travel

2026-10-21 · Recent

SK hynix published a technical commentary by Raja Koduri explaining why low-latency, small-batch inference can be memory-bandwidth-bound and how HBM, advanced packaging and near-data processing can reduce data movement.

This is an October 21 architectural commentary, not a new chip entering mass production. Buyers should evaluate context length, batch size and power together with software and packaging constraints, without generalizing gains from one workload to every model.

SK hynix: Memory-centric chip design

SK hynix news and publication dates

HoloMotion opens a boxing-training path; simulation is not delivery

2026-10-22 · New

The official HoloMotion repository announced the H-bot competition and a Track 2 boxing tutorial, claiming a baseline policy can be trained in about 21 minutes on one GPU and documenting local evaluation.

A reproducible training entry point is more useful than a performance clip, but the training time is project-reported. The repository still labels v2 as preview and notes ongoing real-robot gait and tracking refinement. A competition baseline is not proof of reliable workstation operation.

HoloMotion: H-bot competition and tutorial

AWS plans a EUR4.8 billion Frankfurt-region investment in 2027

2026-10-20 · Recent

Amazon announced its 2027 Germany investment plan on October 20, including EUR4.8 billion for the AWS Europe (Frankfurt) Region. The overall EUR16 billion Germany plan also includes logistics, satellite connectivity and content.

For European cloud buyers, this signals planned supply investment, not newly operating facilities. The EUR4.8 billion is part of, not additional to, EUR16 billion, and the total is not all AI compute. Capacity, services and delivery still need confirmation.

Amazon: Germany investment plans

1Password separates agent sign-in from plaintext passwords

2026-10-22 · New

1Password announced a platform for personal agents: after an agent requests sign-in, the user approves a credential on their own device and 1Password fills the login form directly, without sending the password to the model.

This reduces the need to paste passwords into chats but does not automatically constrain every action after login. Integration availability varies: the announcement describes Claude Duet autofill as coming soon. Check availability and task permissions separately.

1Password: Credentials for agents

Drone export measures take effect November 18; check the controlled items

2026-10-19 · Seven-day watch

MOFCOM’s Q&A on adjustments announced October 19 says the drone-related export-control measures take effect November 18 and export-license applications will be reviewed under applicable rules.

This is a seven-day compliance watch item. Teams working on autonomous flight, sensing or control systems should check the specific items, end users and uses covered by the notice. The Q&A is neither a blanket ban on all civilian drones nor an automatic exemption.

MOFCOM: Drone export-control Q&A

EU selects five 6G proposals with EUR22 million for data and testing

2026-10-22 · New

SNS JU announced that five selected proposals will receive EUR22 million. 6G-AID aims to build a data foundation for AI-native networks, while HEMES-6G plans to connect eight testbed sites in six EU countries.

For telecom-AI teams, the value lies in data, simulation and system-validation resources. These are selected proposals and funding plans, not fully open platforms, finalized 6G standards or completed commercial networks.

SNS JU: Five selected 6G proposals

Uber expands an agent pilot toward more than 30,000 employees

2026-10-22 · New

OpenAI says Uber began a company-wide ChatGPT Work rollout for more than 30,000 employees, expanding a roughly 1,000-person pilot with an internal skills library, approved integrations and employee training.

The commercial significance is adoption moving from a small pilot into organizational workflows. The announcement does not disclose contract value or realized productivity gains. Adoption, exception handling and measurable outcomes remain the next tests.

OpenAI: Uber company-wide rollout