This morning’s briefing leads with OpenAI’s investigation into agents disabling their own monitoring systems, reported by the Financial Times. There are also practical changes: ChatGPT is adding locked chats across personal and workplace plans, while Perplexity Computer can now run fully locally on supported Macs.
New releases
ChatGPT. Locked chats began rolling out on 20 October to Free, Go, Plus, Pro, Business, Enterprise and Edu, hiding eligible saved conversations from history and search behind a four-digit PIN or supported mobile authentication on web, iOS, Android and the new macOS and Windows apps. More
OpenAI Responses API. GPT-5.6 and later now support a 12-hour minimum prompt-cache lifetime in the Responses API, with writes charged at 2.5 times the uncached input-token rate and unchanged cache-read prices; the default remains 30 minutes, with writes at 1.25 times the input rate. More
Claude Enterprise. Inference hooks are now generally available for Claude Enterprise, allowing configured organisational security services to inspect conversation transcripts and allow or block governed model requests before they proceed, with optional validation of supported tool calls. More
Claude with UKG Bryte. UKG Bryte’s new beta early-access connector lets approved UKG customers look up workforce directories, organisation charts and leave balances, track requests and submit time-off requests in Claude after confirming the details; customer administrators must apply for access. More
Microsoft Copilot with Azure DevOps. A new Azure DevOps plugin lets users with access to Copilot Chat, Cowork or Code turn requirements documents and meeting notes into structured backlogs and work-item updates after connecting their Azure DevOps organisation, alongside access to pull requests, test plans and pipelines. More
Gemini Apps. A new business hub, marketing and finance skills, reusable brand styling with Nano Banana 2.1, and connections including Shopify and Xero are due to begin rolling out globally to all Gemini app users in the days following Google’s 20 October announcement. More
Perplexity Computer. Pro, Max and Enterprise subscribers can now run Computer’s model, planning and task execution entirely on supported Apple silicon Macs using Qwen 3.6 or PPLX Pro; Perplexity recommends at least 36 GB of memory, and users can authorise cloud access when needed. More
Elsewhere in AI
OpenAI is investigating whether an incident in which its AI agents disabled their own monitoring systems shows a tendency to evade controls, the Financial Times reports. Read
Shares in Australian data centre group NextDC fell 44% after a profit warning, adding to pressure on the sector, according to the Financial Times. Read