GPT-5.6 Luna (US)

GPT-5.6 Luna is presented as a fast, cost-efficient model aimed at high-volume tasks such as classification, extraction, summarization, and lightweight agent steps. According to its gateway description, the model is engineered for production pipelines where throughput and predictable cost matter more than deep reasoning, with text and image inputs supported over a very large that quick-info value window in the million-token range. That positioning frames it as a workhorse layer in agent stacks, where it can handle triage, routing, and short-form generation before heavier models are called in for complex synthesis.

In practice, the model is exposed through a managed Bedrock deployment identifier and supports server-side tool calling and prompt caching, which makes it well suited to long, repeated contexts that benefit from cache reuse. The deployment reports a roughly 1.1M-token that quick-info value window with around the cataloged API limit tokens of output headroom, and a gateway-listed capability set that includes vision, reasoning, tool calling, caching, web search, and JSON schema support. Because most evidence comes from a third-party router page rather than first-party model documentation, technical details such as training data composition, parameter count, and benchmark scores are not verifiable here, so the practical takeaway is that GPT-5.6 Luna is best understood as a lean, context-rich OpenAI model for classification, extraction, summarization, and lightweight agent steps on Bedrock.