The cheapest Clay credit is the one you never spend. Since the March 2026 split into Data Credits (provider lookups) and Actions (workflow steps, AI columns, HTTP requests, CRM pushes), the optimization calculus changed: failed lookups now cost nothing, but unfiltered rows and over-provisioned AI still burn both meters. The four moves that actually move the bill are, in order: connect your own provider API keys (0 Data Credits on those lookups, 70–80% typical reduction), filter and dedupe rows before any paid step runs, wrap every expensive column in "only run if empty", and match the Claygent model to the task (Helium 1 credit, Neon 2, Argon 3). A 10,000-row monthly table on Clay-managed keys can run six figures of credits a year; the same table on your own keys drops the data side toward zero and leaves only Actions.
- BYOK is the single biggest lever. Connect your own Apollo, Hunter, Clearbit, or LLM keys via Settings → Connections and those lookups cost 0 Clay Data Credits. Actions still apply, but data spend collapses 70–80%.
- Filter before you enrich. Remove personal email domains, duplicates, and rows missing required fields before the first paid column. A row that should never have been in the table is the most expensive row you have.
- Waterfall order is a cost control. Put your highest-coverage provider first so fewer rows fall through to expensive fallbacks. Since March 2026, failed lookups cost neither credit nor Action.
- Match Claygent to the job. Helium (1 credit/row) for high-volume lookups, Neon (2) for qualification, Argon (3) for multi-step research. Running Argon where Helium would do is pure waste.
- Replace AI columns with formulas wherever a JS formula answers the question. Formulas are free and instant; Claygent calls are not.
- Watch the credit dashboard weekly to find the two or three columns eating the budget, and model cost per qualified lead before scaling.
- Why the March 2026 split matters
- Move 1: connect your own provider keys (BYOK)
- Move 2: filter before you enrich
- Move 3: waterfall order and conditional logic
- Move 4: match Claygent to the task
- Price your waterfall with and without BYOK
- Failure modes that quietly drain the balance
- Where each tactic lands on the bill
- What we actually run
- FAQ
- Related Cluster Intelligence
Why the March 2026 split matters
Before March 2026, Clay billed a single credit per enrichment. After the split, every enrichment draws from two separate meters that deplete independently:
- Data Credits pay the external providers in Clay's marketplace — email lookups, phone numbers, firmographics, and AI model calls routed through Clay's negotiated rates.
- Actions pay Clay for running the workflow itself — each enrichment step, AI column, HTTP request, or CRM push counts as one Action regardless of weight.
That distinction is the whole optimization story, because the two meters have different free levers. Data Credits drop to zero the moment you bring your own provider key; Actions do not. So a cost-reduction plan has to attack the Data Credit side with BYOK and the Action side with fewer, conditional steps — not one generic "use less Clay" instruction.
Two meters, two fixes. BYOK zeroes the data side. Conditional logic and filtering shrink the action side. Treating them as one bill is why most teams over-spend.
Move 1: connect your own provider keys (BYOK)
This is the highest-impact change and it is one-time setup. In Clay, go to Settings → Connections → Add Connection and paste your provider API key. Every enrichment using that provider then consumes 0 Clay Data Credits — you pay the provider directly at their published rate, and Clay's markup disappears.
The savings are large because Clay's convenience rate is a markup over the provider's own price. Representative Clay-managed costs cited across 2026 guides: Apollo ~2 credits/lookup, Hunter ~2, Prospeo ~2, Clearbit 2–5, People Data Labs ~3, ZoomInfo 5–13. With your own key, every one of those becomes 0 credits. LLM keys are even more dramatic — routing Claygent through your own Anthropic or OpenAI key drops per-row AI spend from 1–5 cents to fractions of a cent, an 80–95% cut on the AI side.
The catch to state plainly: BYOK removes the Data Credit charge, not the Action. Clay still counts the workflow execution. So your bill does not go to zero — but for data-heavy and AI-heavy tables, the data side is usually the larger driver, which is why 70–80% total reductions are typical.
Move 2: filter before you enrich
The most expensive credit is the one spent on a row that should never have been in the table. Before any paid column runs, build disqualification logic that flags non-ICP rows: company size, region, industry, tech stack. If a row fails the filter, it never touches a provider.
Concretely, before enrichment:
- Deduplicate by name + domain.
- Strip personal email domains (gmail, yahoo, outlook, icloud, hotmail).
- Drop rows missing required fields (first name, last name, company domain).
- Qualify the company before pulling its people — enriching 1,000 contacts at companies you would never sell to is the fastest way to torch a balance.
A team blowing through 100k credits a month is usually not doing anything advanced. They are running expensive providers on leads that were never going to convert.
Move 3: waterfall order and conditional logic
A waterfall queries providers in priority order and stops at the first verified hit. Put your highest-coverage, lowest-cost provider first so fewer rows fall through to expensive fallbacks. A 5-provider waterfall at 10–15 credits/row can be cut to a 2-provider waterfall at ~4 credits/row with only a single-digit coverage loss — roughly two-thirds of the credit spend, gone.
Then wrap every expensive column in a condition: "only run if email is empty", "only run if company passed the ICP check", "only run if the previous step returned nothing". This stops Clay re-enriching rows that already have data and stops AI columns firing on rows that failed earlier filters.
Stop-at-first-hit is the gift March 2026 gave you: a failed lookup now costs nothing. Make your waterfall wide enough to cover, but order it so the cheap provider answers first.
Move 4: match Claygent to the task
Claygent bills per model. Helium at one credit per row handles high-volume lookups; Neon at two suits qualification and structured answers; Argon at three is for genuine multi-step research. Frontier models such as GPT-5.1 and Claude 4.6 Sonnet bill on actual token use. Matching the model to the task is one of the easiest spend cuts — running Argon where Helium would do is paying 3x for the same answer.
Wherever a formula answers the question, use a JS formula column instead of an AI column. Half the Claygent calls in a typical workspace could be a formula: free, instant, and reproducible.
Price your waterfall with and without BYOK
This calculator sizes your monthly credit consumption and estimated data spend for a table, toggling how much of the waterfall runs on your own keys.
Failure modes that quietly drain the balance
Enriching before filtering
Running paid columns on unqualified, duplicated, or personal-domain rows is the most common leak. Every one of those rows costs Actions and possibly credits for nothing.
Wide waterfalls with no ordering
A 5-provider cascade where the expensive provider answers first burns credits on rows the cheap provider would have resolved. Order by coverage and cost, not by habit.
Argon for everything
Defaulting every AI column to the most capable (and most expensive) Claygent model triples the credit cost of lookups that Helium would have answered identically.
AI columns that should be formulas
Any transformation a JS formula can do is free. Replacing even a few Claygent columns with formulas removes their per-row cost entirely.
Where each tactic lands on the bill
| Tactic | Meter it attacks | Typical impact |
|---|---|---|
| Connect own provider keys (BYOK) | Data Credits | 0 credits on covered lookups; 70–80% total reduction |
| Filter + dedupe before enrich | Both | Removes wasted rows entirely; often 20–40% fewer rows processed |
| Waterfall order + stop-at-first-hit | Data Credits | Fewer fall-throughs to costly providers; ~third of spend on wide waterfalls |
| Conditional "only run if empty" | Actions + Data Credits | Stops re-enriching rows that already have data |
| Match Claygent model to task | Data Credits (AI) | Helium/Neon/Argon at 1/2/3 credits; right-sizing cuts AI spend |
| Replace AI columns with formulas | Actions + Data Credits | Free and instant where a formula suffices |