Gemini 3.8 Flash Stable

This model runs on a channel we operate directly, separate from the shared Token Codex pool.

• You get the model you asked for. Requests are never quietly served by a smaller or older model.
• No queueing behind pool contention. Capacity on this lane is not shared with contributed accounts, so a busy pool does not slow you down.
• Prompt caching and extended thinking pass through unchanged, so cached turns stay cheap and reasoning behaves the same as on the provider's own API.

This is a single direct connection, not a multi-provider pool: it trades the pool's failover for predictable behaviour. If this channel is down, requests to this model fail rather than falling back.

Measured on our side. A dash means the window has no measurement yet — never a failure.