AI · Costs and operationsOctober 18, 20264 min · updated October 18, 2026
Vercel groups AI Gateway logs by session: measure the full conversation
Daillac research
Vercel — AI Gateway logs now show session affinity, October 17D
Vercel announced a Sessions view for AI Gateway logs on October 17, 2026. It groups model requests from a conversation, making it easier to investigate spending over several turns. For a business pilot, the next question is how much of that spending produced a usable result.
What appears in the new view
The announcement describes session-level cost, models, providers and cached input, with a path through the conversation’s turns. Filters help isolate sessions by criteria such as model, provider or cost. These are observations about requests passing through AI Gateway; they do not establish the profitability of a workflow.
The session-affinity documentation explains how successive turns can remain with the same provider. A stable provider may reuse more cached input. Applications can supply a consistent session identifier to group a conversation; the way a business case maps to that conversation still needs to be defined by the application team.
Keep the business case identifier separate
A customer request may include an initial conversation, a retry after a failed tool call and a human correction. If each part creates a new session, reporting only the first session understates the work required. Keep an internal case identifier that allows authorized staff to reconcile these steps without exposing personal details in a public label.
TableChoose the unit that answers your question · shareable block
| Useful for | What remains outside the measure | |
|---|---|---|
| Individual model call | Finding an expensive request or unexpected model. | Other turns, retries, tools and the final business outcome. |
| Conversation session | Reviewing a multi-turn interaction. | Work split across other sessions and employee intervention. |
| Completed business case | Comparing the cost of an accepted result. | Needs application records and a definition of acceptance. |
A calculation to prepare before changing models
Use an illustrative period with 100 submitted cases, 80 accepted outcomes and CAD 100 in the costs you have chosen to include. That is CAD 1.25 per accepted outcome, compared with CAD 1 per submitted case. These are fictional planning values, not Vercel prices or a Daillac customer result.
State the cost boundary. Model usage can be one line; paid external tools, infrastructure and employee correction time can be separate lines. Keep currencies and the observation period consistent. If labour is estimated rather than recorded, label that estimate so it is not mistaken for a billed amount.
TableProposed pilot record · shareable block
| Record alongside spending | Decision it can support | |
|---|---|---|
| Accepted outcome | What qualifies as a usable answer or completed action. | Whether a less expensive route still does the required work. |
| Retry | Reason and case identifier. | Whether an integration failure is driving repeated calls. |
| Human correction | Time or clearly labelled estimate. | Whether lower model cost transfers work to employees. |
| Long session | Purpose and eventual outcome. | Whether length is necessary or caused by a repeated failure. |
Compare complete runs under the same conditions
Take a small set of synthetic requests representing routine work and difficult exceptions. Keep the expected outcome and tool permissions fixed while comparing configurations. A cheaper model is worth keeping when the required outcomes still pass; a shorter conversation that omits an important check is a different result.
- Review expensive sessions with their final outcome.
- Separate useful follow-up questions from loops caused by missing data.
- Document model or routing changes so period comparisons remain understandable.
- Restrict log access and review which request data is retained.
We would use these observations to refine a pilot before increasing volume. Our AI agent budget guide separates initial setup from ongoing operation. Session logs add evidence to that operating review; application records are still needed to count work completed.
IllustrationEditorial still life of branching paper strips with a stopwatch and calculator representing conversation-level costs · shareable block
Questions fréquentes
Do these logs include every business cost?+
No. The announcement concerns AI Gateway request and session observations. Tool costs, other infrastructure and employee work require their own records.
Does session affinity guarantee savings?+
No. Cache use and billing depend on the requests, provider and current pricing. Compare observed costs for equivalent outcomes.
Should a case ID contain a customer’s name?+
Use an identifier appropriate to your data policy and access controls. Staff can reconcile it with the business system when authorized.
What should we measure first?+
Define an accepted outcome and count it alongside the costs, retries and correction effort included in your pilot.
Scope an AI pilot with usable cost evidence
Choose one process and define what a completed case means before comparing models or increasing volume.
Sources & method
Primary-source review dated October 17, 2026. Vendor statements are attributed; practical recommendations and hypothetical examples are Daillac editorial analysis.
News feed
October 18, 2026
Security · Automation
n8n security update: check your version and execution logs
October 18, 2026
Security · Application testing
AWS reports 33 of 37 security-test successes: what evidence to ask for
October 12, 2026
Development and hosting