Models, effort, cost and insightsGrid Console does not call models. It passes a model and an effort level to the active provider per card, and reads usage back from that provider’s session data.Picking a modelAt capture, and again per stage if you want. The default comes from your profile; the advice comes from the card description. The list itself comes from the active provider, so the names below are the ones the bundled Claude provider contributes. Another provider fills the same picker with its own models, and Grid Console only needs to know which of them is the capable one, the everyday one and the cheap one.MODEL · CLAUDE PROVIDERWHENFableThe most capable, and the most expensive. Complex, cross-cutting work: migrations, refactors that span repos.OpusVery complex tasks, at lower cost than Fable.SonnetEveryday work. The right default for most cards.HaikuSmall, mechanical changes and fast explanations in the code view.
EffortHow much thinking the provider is asked to spend: low, medium, high, xhigh, max, and ultracode. Ultracode is the deliberate outlier: maximum reasoning with self-review between steps, for work you would otherwise pair on. It is slow and expensive, and Grid Console will tell you when it is not warranted.Effort is a Grid Console concept, not a provider one. A provider plugin declares which levels it supports and how each maps onto its own knobs: a thinking budget, a reasoning parameter, or a different model entirely. Levels a provider does not support are hidden rather than silently downgraded, and the one that stands in for ultracode is named in the picker.The advice lineRight after you type the description, Grid Console suggests a model and effort and says why in one line. What it suggests depends on your profile: your subscription, your token budget, and whether you have said you would rather spend tokens than time. It is written against the roles a provider declares (capable, everyday, cheap), not against model names, so the advice keeps working when you switch providers. With plenty of headroom it reaches for the capable model, Fable on the Claude provider; on a tight budget it picks the everyday one and says so.InsightsTokens and cost per card and per request, by model, by project, over 30 days, so a surprising bill has a cause you can find. Per card you see every request with its model, effort, tokens in and out, and cost; per project, spend over time and which stage burns it.The numbers come from the active provider reporting usage back. A provider plugin that does not report usage leaves Insights empty for its sessions, and the panel says which provider is silent rather than guessing.