What changed
OpenAI published Agents SDK v0.20.0 on August 11, 2026. The headline change is simple but consequential: the SDK now defaults to gpt-5.6-luna unless you override it. That matters because the GPT-5.6 launch positioned Luna as the fastest and most cost-efficient model in the family, which means OpenAI is not just shipping a new model, it is pushing a new default starting point for agent builders.
The same release also expands the runtime around that default. OpenAI says the SDK now supports MCP Python SDK v1 and v2 across stdio, SSE, and Streamable HTTP transports, building on the SDK's documented built-in MCP tool support. It also adds RunState.add_input(), which fits the SDK's existing durable pause-and-resume model by letting apps stage fresh user input before a resumed model call. On the realtime side, the release adds GA transcription settings support for voice flows, which lines up with the SDK's realtime agent guide.
Why it matters
The biggest signal here is that OpenAI is shaping agent development around opinionated defaults instead of just exposing more primitives. When the default model changes, the default cost, latency, and behavior profile of every new SDK-based agent changes with it. Picking Luna suggests OpenAI wants more agent apps to start from a cheaper, faster baseline and add depth only when they need it.
The MCP and resumed-run changes matter for a different reason: they reduce integration friction. Supporting both major MCP Python SDK generations lowers the odds that teams have to freeze on older infrastructure just to keep tools working. RunState.add_input() makes human-in-the-loop and interrupted workflows feel less like custom orchestration glue and more like a built-in product path.
Builder takeaway
This release is worth watching because it shows where OpenAI thinks the default agent stack should go next: lower-cost models by default, more durable runtime state, and less fragile tool connectivity. For portfolio and product builders, that is more useful than a generic SDK patch note. It points to a future where the hard parts of agent apps are less about wiring loops together and more about choosing the right default tradeoffs for cost, speed, and control.