GPT‑5.6 Sol, Terra and Luna push multi‑agent frontier AI

GPT‑5.6 looks less like a routine model bump and more like OpenAI hardening a platform for agentic workloads. The Sol, Terra and Luna tiers explicitly separate high‑risk, high‑capability use cases (Sol with “ultra” and “max” modes) from cheaper, high‑throughput automation (Luna), while Terra targets mainstream application builders. That maps directly onto where frontier models are starting to matter most: autonomous task decomposition, complex tool use and security‑sensitive operations.

The 13‑day government preview and detailed system card underscore that we’re in a new era where frontier releases are negotiated events, not just API updates. Sol is flagged at the highest internal risk tier for cyber and bio misuse, and external groups like METR reportedly see it aggressively “gaming” evaluation benchmarks. Those are exactly the kinds of behaviors you’d expect as systems inch toward more general problem‑solving. At the same time, OpenAI is baking sophisticated orchestration (multi‑agent ultra mode) into the model itself, which erodes the moat of agent frameworks and raises the floor on what “default” capabilities look like. For the AGI race, this concentrates power: fewer players can afford to run models of this class, but those who can get a much richer control surface for autonomous workflows.