GPT-5.6's agent API: the model writes your tool-calling loop
The Responses API now runs orchestration, subagents and persisted reasoning server-side — one team cut input tokens 21%, another lifted an agent score 13→38%.
Copy markdown
The model writes the orchestration, not you
Programmatic tool calling lets GPT-5.6 emit JavaScript that runs your tools in a sandbox — loops, branches, aggregation — returning once instead of N serial round-trips. A financial-research team booked 21% fewer input tokens at the same quality, since intermediate data never re-enters the context.
Multi-agent fan-out, no framework
Native multi-agent orchestration coordinates a primary agent and parallel subagents straight from the Responses API — no LangGraph or Swarm glue. One builder ran six concurrent specs "without the quality falling apart," getting fan-out patterns for free.
Retained reasoning is a real long-horizon jump
Reasoning now persists across turns with native compaction. On ARC-AGI-3 that lifted a run from 13.3% to 38.3% while spending ~6x fewer output tokens — no model swap — so long agent loops stay coherent and cheaper.
Prompt cache: 30-minute floor, deterministic breakpoints
The minimum prompt-cache TTL is now 30 minutes with deterministic breakpoints you place in-context. One team cut uncached input 28% on a 29K-token prompt using breakpoints plus workspace-specific keys — straight savings on repeat calls.