OpenAI's Responses API adds async tools and mid-turn steering
Astra's launch shipped agent-native API primitives — async tools, live steering, mid-turn reasoning control — but they're Responses-API and Astra-only.
Copy markdown
Async tools: the model stops waiting
Set `async: true` on a function tool and Astra keeps reasoning while your slow database or API call runs in the background — you return the result later by its `call_id`. It kills the block-and-wait stall that idles agent loops. GPT-6 Astra and later only; your app still executes the tool, OpenAI doesn't.
Mid-turn steering: redirect a live run
Over a WebSocket, send `response.steer` to inject new instructions while Astra is still reasoning — no cancel-and-restart. It won't rewrite output already streamed or kill tools mid-flight, and it's Astra-only (GPT-5.6 and earlier can't). Practical for course-correcting a long autonomous task the moment you see it drift.
Dial reasoning effort mid-conversation
You can now raise or lower reasoning effort partway through a conversation while keeping cached prefixes intact — pay for deep thinking only on the hard turns. The trade-off: Astra drops the `none` effort level entirely and locks temperature and top_p.
The catch: Responses API, Astra-only
All three primitives require the Responses API — tool calling on Astra won't run through Chat Completions at all — and none of it back-ports to older models. Adopting them means committing an agent to Astra and OpenAI's newer surface.
Elsewhere: Gemini cuts long-video tokens 88%
Google's agentic video understanding, now across Gemini 3.7/3.6 Flash and 3.5 Flash-Lite, lets the model pull transcripts, frames, or audio on demand rather than ingesting everything — up to 88% fewer tokens on long clips. A far cheaper path to hour-long-video pipelines.