OpenAI wires Decisions into GPT-Live: voice agents act mid-sentence
Native delegation kills the relay server you had to hand-build; Decisions runs ~10x faster than Responses at $0.10/1M tokens. Plus Claude Code 2.1.301.
Copy markdown
Native delegation drops the middleman server
GPT-Live can now call the Decisions API itself as transcript deltas arrive, instead of your backend reading state, hitting /v1/decisions, and piping the answer back. Create the session with native delegation and it's one less service to run and keep in sync.
Act while the caller is still talking
A decision fires on input or output transcript deltas, so your app can start an action — pull an order, check availability — before the sentence ends, then speak the result via session.commentary.append. Full-duplex voice agents stop feeling turn-based.
The price-and-latency math
Decisions input runs $0.10 per 1M tokens on gpt-6-luna with no output or cache charges; GPT-Live voice stays $0.05/min. OpenAI claims Decisions is about 10x faster than routing the same choice through the Responses API — no p95 published, so measure it yourself.
Claude Code 2.1.301's quiet quality-of-life drop
Switching to a smaller-context model now auto-summarizes an over-long chat first instead of erroring, and the unattended retry watchdog won't sit waiting out a usage limit that resets days away. Vim mode gains % bracket-jump; VS Code renders LaTeX in replies.