Vercel’s AI Gateway logs now group related requests into sessions, showing spending, providers and cached inputs across a conversation. Developers can inspect individual turns and see whether routing kept them with the same provider. Explicit session identifiers let applications choose their own grouping; otherwise the gateway uses the conversation’s opening. The new view concerns request history, rather than the earlier figures on model adoption.
Vercel has added a Sessions view to the logs of AI Gateway, its service for routing model requests. Related requests are grouped together, including work sent by Codex, Cursor or Claude Code. Each session shows spending, models, providers and input served from cache across its turns. Opening it reveals duration, cache hit rate and the route taken by individual requests.[1]
Session affinity sends each later turn to the provider that served the preceding one, with the aim of reusing prompt material already in its cache. The new view shows this routing behavior alongside cache use. Filters cover client, model, provider, authentication and cost. Single-turn sessions remain hidden until the minimum-turn filter is cleared.[1]
Text and chat requests use session affinity by default. An application can choose the identifier used for its sessions through the AI software development kit, or SDK. Direct requests can instead carry the x-session-id header. Each request in a conversation must carry the same chosen identifier. Without one, the gateway groups requests using a fingerprint of the conversation’s opening. The primary release document is the sole source for this change; its observations cover requests passing through AI Gateway.[1]
