GPT-5.6 preview adds cost and cache rules

OpenAI says GPT-5.6 Sol, Terra, and Luna are in limited preview for selected API and Codex partners, with published token prices and cache-write/cache-read rules. Why it matters: The operational signal is access and routing, not broad adoption: stronger agent models may arrive with gated availability, different Codex approval paths, and pricing that rewards deliberate prompt caching.

Try this: Record the preview access limits, per-token prices, and cache behavior in a model-routing table, but avoid building production paths that assume self-service GPT-5.6 access.