Claygent Adds Open-Weight Models: Kimi K2.6 and GLM 5.2

On July 16, Clay quietly added two open-weight models to Claygent: Kimi K2.6 and GLM 5.2. If you only use Claygent for the occasional lookup, this won't change your week. But if you run AI research columns across thousands of rows a month, model choice is a real budget line — and this is Clay giving you a cheaper dial to turn.

What Clay shipped

From Clay's changelog: the release is in beta, and the two models are available as options in any Claygent or "Use AI" column. Clay's pitch is that you can now run "frontier-quality GTM research at lower cost."

Quick context if the names are unfamiliar: Kimi (from Moonshot AI) and GLM (from Z.ai) are two of the most capable open-weight model families available right now. "Open-weight" means the model's weights are published, so they're cheap to serve compared to closed frontier models like Claude or GPT. The old tradeoff was that open models lagged badly on quality. That gap has narrowed a lot — especially for the kind of structured research work Claygent actually does.

The changelog doesn't publish specific credit pricing for the new models, so check the model dropdown in your own workspace before you plan around exact numbers.

Why this matters for outbound teams

Here's the thing about enrichment costs: they scale with list size, not deal size. A 10,000-row TAM build with three Claygent research columns is 30,000 AI runs. At that volume, the per-row cost of the model doing the research matters more than almost any other setting in your table.

And most Claygent tasks are not hard reasoning problems. "Does this company do commercial or residential work?" "How many locations do they have?" "Find the pricing page and tell me if they sell to enterprises." That's structured reading and extraction. You don't need the most expensive model on the market to do it well — you need one that follows instructions and doesn't hallucinate when the answer isn't there.

Cheaper capable models change the math in two directions. Either your current research bill drops, or — more interesting to us — you run research you were previously rationing. Every operator we know has qualification questions they'd love to run table-wide but skip because the credit burn wasn't worth it. This is how those come back.

How we'd use it

We run Clay daily for client list building, so here's the practical playbook:

• A/B it on 100 rows before you switch anything. Duplicate an existing Claygent column, point the copy at Kimi K2.6 or GLM 5.2, and run both on the same sample. Compare fill rates and spot-check 20 answers by hand. Ten minutes of checking beats discovering a quality drop after a 5,000-row run.
• Tier your columns. Use the open-weight models for binary qualifiers, firmographic lookups, and website summaries. Keep a frontier model on the columns where tone and judgment show up in the output — like personalization lines that go straight into email copy.
• Re-run the research you've been skipping. Got a qualification question you only ran on high-priority segments because of cost? Run it across the whole table with the cheaper model and let it resurface accounts you wrote off.
• Watch your fill rates after switching. Clay shipped function observability the week before this — rows processed, credits per row, output fill rates. Use it. If fill rate dips after a model switch, that's your signal to move that column back up a tier.

The mistake to avoid: switching every column to the cheapest model on day one because the line goes down. Research that's 15% wrong is more expensive than research that costs 2x, because bad data flows downstream into segmentation, copy, and calls. Test, tier, then commit.

FAQ

What are open-weight AI models?

Models whose trained weights are publicly released, so any provider can host and serve them. That competition makes them significantly cheaper to run than closed models. Kimi (Moonshot AI) and GLM (Z.ai) are two of the leading open-weight families.

Are Kimi K2.6 and GLM 5.2 as good as Claude or GPT for GTM research?

For structured research — extraction, classification, qualification — recent open-weight models get close enough that many tasks won't show a difference. For nuanced writing and judgment calls, frontier closed models still tend to win. Test on your own columns; your data will answer this faster than any benchmark.

Do the new Claygent models cost fewer Clay credits?

Lower cost is the entire positioning of the release, but the changelog entry doesn't list specific credit pricing. Check the model selector in your Claygent column for current rates — and remember the feature is in beta, so details may shift.