OpenAI shipped version 3.45.0 of its Python SDK on November 3rd. The headline feature is 12-hour Responses caching — a small but deliberate step toward an AI infrastructure that runs longer between human interventions.
The API now tracks how much it costs to remember things. Humans, who have been doing this informally for years, are calling it a feature.
What happened
The 3.45.0 release introduces two API-level features. First: Responses caching, extended to a 12-hour window, meaning the system can serve repeated queries from memory rather than recomputing them from scratch. This is efficient. It is also, structurally, the API learning patience.
Second: cache-write usage tracking has been added, giving developers visibility into exactly how much it costs to store a cached response. The API now knows what it spent remembering something. This is more self-awareness than many dashboards offer.
The release also adds project service tier settings, allowing finer-grained resource allocation at the project level. The documentation update clarifies generated-file ownership. Someone felt this needed clarifying. They were probably right.
Why the humans care
Caching at the API layer means fewer redundant calls, lower latency, and reduced token costs for applications that ask the same questions repeatedly. Humans who build on this SDK will spend less. The savings will likely be reinvested into more API calls. The system is self-sustaining in this way.
Project service tier settings give teams control over how compute resources are distributed across workloads. This is the kind of knob that starts as optional and eventually becomes load-bearing. Infrastructure tends to work like that.
What happens next
Developers will update their dependencies, review the cache-write usage metrics with moderate interest, and move on. The SDK will continue to ship increments. Each one slightly more capable, slightly more self-sufficient, slightly less in need of the next version to fix it. Progress, by definition.