Clay's Auto-Mode Saves Clay 47% on Inference. Your Credits Depend on the Row

Clay made auto-mode the default for new Claygent columns on October 27. In its announcement, Clay said the mode picks a model for each row, sending simple lookups to fast models and complex research to the most capable ones. On production traffic, Clay said, auto-mode cost 47% less than routing every task to its most capable model, while keeping what it called frontier quality. Existing Claygent columns keep the model they already use, and any column can still be pinned to one model.

Applied Compute, which trained the router with Clay, describes the 47% in its case study as a cut in Claygent's inference costs. It was measured in a live A/B test against sending every request to GPT-5.6 Sol, and the case study's routing examples use GPT-5.4 Nano and Mini as the cheaper picks. Offline, the router saved 48.2% at a fixed quality level. Both figures come from the two companies that built the router, and neither published the mix of tasks behind them.

Clay passes reasoning-model costs through and says less about light models

Clay bills in two units: actions for platform work, and data credits for data and AI. Its AI pricing guide says auto-mode is variable-priced, charged "at the rate of whichever model actually ran that task." A row the router sends to a lighter model costs the customer fewer credits, and Clay's announcement says as much. Every AI prompt still counts as one action, so the router changes the data-credit side of the bill and leaves the action count alone.

How much of the 47% reaches a customer's credit balance depends on where the rows land. Clay charges 0% markup on its variable-priced reasoning models. It made this commitment when it rebuilt its pricing on March 11. Lightweight models, about 80% of its catalog, kept flat credit prices per task. Its public price table doesn't list the GPT-5.4 Nano and Mini models in Applied Compute's examples. In March, Clay also said flat rates had left customers paying for other customers' heavier prompts. Our read is that a flat price is an average across many customers' rows. Clay hasn't published how its flat prices for light models compare with what those models cost it to run.

Auto-mode won't lower every team's bill either. Clay's post says harder rows go to stronger models and cost more. A team that already pinned a cheap model on a table of simple lookups was never paying the all-Sol baseline behind the 47%. It could pay more on auto if the router reads its rows as hard. Clay tells customers to test auto-mode on five to ten rows before running a large table.

Enrichment vendors now differ on who picks the model

Clay keeps the routing decision inside its product. ZoomInfo is pushing the other way. Henry Schuck, the founder and CEO of ZoomInfo, which competes with Clay, posted on October 21 that he runs enrichment from Claude and Codex through ZoomInfo's GTM.AI command-line tool. In that setup the customer's own AI tool picks the model and pays for it, and the data vendor charges for the data. ZoomInfo already sells its data into other companies' agents, most recently through Exa. The model margin sits with Clay in one design and with the customer's AI provider in the other.

What this means for RevOps leads who own the enrichment budget

Before you move existing tables to auto-mode, rebuild one as a new column beside its pinned version and run both on the same sample. Compare data credits per row and answer quality. Keep the pinned column wherever a score feeds a routing rule and consistency matters more than cost. Your old columns won't switch on their own, and duplicates keep the old model too. So the saving only reaches columns you create or switch yourself.

Then ask Clay for two numbers on your own account: the per-row credit price of each model auto-mode can route to, and the share of your rows each model handled last month. Clay shows the realized cost per row on variable-priced runs, which gives you something to check its answer against.

Put the same questions to every enrichment vendor with an AI agent in its product. Ask which models it routes to, who picks, and whether a cheaper model shows up as a lower charge on your invoice or as the vendor's margin. A vendor that bills a flat rate per AI task keeps any routing saving until it cuts the rate, and Clay has published only part of that answer for its own rows.

By Business of GTM