Browse LLM Gateway models

Models are listed as maker/model. The part after the slash is the exact model value to pass in a request. The Gateway does not accept the maker prefix.

Rates are for global routing. In-region US and EU endpoints are 10% higher, passed through from provider pricing with no AssemblyAI upcharge. See pricing for the full rate card.

Cache read is the rate for tokens served from the prompt cache; cache write is the rate to store them. An em dash means the model has no cache tier. See prompt caching.

Recommended marks the 8 best picks for each transcript task in the Gateway's own benchmark: dictation, information extraction, summarization and text correction, each measured on transcripts from 30 seconds to an hour long. A model must average at least 85 out of 100 on the task to be in the running; the 8 that clear that bar with the lowest cost per quality point get the label. An em dash means the model is not in the top 8 for any task, or has not been benchmarked yet. The benchmark is in beta.

Providers are where a request is served from — Claude models run on Bedrock, Gemini on Vertex, and Qwen3.5 4B Fast on AssemblyAI's own GPUs. This table is generated from the Gateway's own Models endpoint, so it always matches what the API accepts. The docs roster adds quality scores and measured latency.