What changed
OpenAI shipped openai-agents-python v0.18.0 on July 7, and the headline change is small but meaningful: RealtimeAgent now defaults to gpt-realtime-2.1.
That change is consistent with the SDK's own realtime documentation. In the Realtime agents quickstart, OpenAI describes realtime agents as server-side, low-latency agents built on the Realtime API and explicitly says new realtime agents should start with gpt-realtime-2.1.
OpenAI has also given the model a dedicated GPT-Realtime-2.1 model page, with a named model entry, snapshots, and published rate limits. That makes the default-switch feel less like a quiet SDK tweak and more like a clearer product recommendation.
Why it matters
Defaults are strategy. Most teams do not spend much time debating which realtime model to wire into a first voice-agent prototype; they follow the path the framework makes easiest.
By moving the default to gpt-realtime-2.1, OpenAI is reducing setup friction for builders who want voice-first agents with the rest of the Agents SDK stack: handoffs, tools, sessions, tracing, and guardrails. It also lowers the odds that teams start on an older realtime configuration and have to migrate later.
This is useful for product builders because it turns realtime voice from a special-case integration into a more standard agent entry point. The SDK is effectively saying that voice agents are no longer an edge demo. They are part of the mainline workflow.
The takeaway
The angle is not that OpenAI launched a flashy new consumer feature today. It is that OpenAI is making a concrete SDK-level choice about what builders should treat as the standard realtime starting point.
That is usually how platforms signal maturity: not just by releasing a model, but by promoting it into the default path inside the framework developers actually use.