OpenAI added organization-level restrictions on the service tiers projects may use on October 21. The previous day it added a 12-hour prompt-cache option to the Responses API. Shared instructions, tools and context can remain cached across a workday, but a 12-hour cache write costs 2.5 times the normal input-token price. OpenAI changelog
The useful calculation is not simply “longer cache equals cheaper.” It depends on how often a stable prefix is reused after the default cache would have expired. A platform team should compare cached reads per write, separate stable policy/tool prefixes from fast-changing conversation state, and keep the longer TTL only where reuse pays for the more expensive write.
Anthropic’s October 21 API version came with new major SDK versions and breaking changes independent of the API version; the SDKs also move from the MIT license to Apache 2.0. Its October 20 controls let administrators disable web fetch or apply organization-wide allowed and blocked domains. Redirect destinations must pass both organization and agent-level rules. Anthropic release notes
That redirect rule is the important security detail. Checking only the requested URL leaves a policy gap if an allowed page redirects to a disallowed destination. The broader lesson is release discipline: pin SDK majors and API versions separately, then test tool failures, redirect handling and enforcement webhooks before rolling forward.