GoHighLevel Conversation AI Now Runs on Claude or Gemini — The Setting That Matters Isn't the Model

Disclosure: I earn a commission if you sign up for GoHighLevel through links on this page. How that works is in the editorial policy.

Everyone’s going to spend this week arguing about which model is “smartest.” Wrong question. GoHighLevel Conversation AI can now run on Claude or Gemini instead of only OpenAI, and the setting that actually protects your leads at 11pm is the one most people will scroll straight past — the fallback model.

What follows is a walkthrough of what HighLevel’s release notes say the feature does and how you’d set it up. So yes, we’ll cover the model dropdown. But stay with me to the fallback section, because that’s the part of this release you’ll thank yourself for in three months.

Your chatbot just stopped being an OpenAI product

Here’s what changed, straight from HighLevel’s Release Radar for the week of October 19, 2026: you can now “select Anthropic Claude or Google Gemini models to power your Conversation AI agents.” Until now, those agents ran on OpenAI models only. The choice is made per agent, “directly from the agent builder” — so your booking bot and your support bot don’t have to run on the same brain.

The same post says the underlying conversational engine was also upgraded to handle multi-part requests, enforce workflow restrictions and process relative dates — the “can I come in next Tuesday?” problem, basically. That’s a separate improvement from model choice, and worth retesting your agent for on its own.

The setup, per the Release Radar, is three moves:

1. Go to Labs and turn on “Claude & Gemini Model Support.”
2. Go to Conversation AI, edit your agent, and pick a Claude or Gemini model from the model dropdown.
3. Optionally configure a secondary model as a fallback, then save and test.

The post says there are “seven new models,” and names four of them: Claude Opus 5.5, Sonnet 5.5, Gemini 3.8 Flash and Gemini 3.5 Flash Lite. I’m not going to guess the other three. Open the dropdown in your own account — that’s the list that counts.

If you haven’t built an agent yet, start with the Conversation AI setup walkthrough and come back. Model choice is the last thing to tune on a bot, not the first. The prompt, the knowledge base and the booking logic are where most of the work lives.

The model dropdown is a tone decision, not an IQ test

Here’s the thing nobody will say in the launch threads. For a Conversation AI agent answering “how much is a cleaning?” and “can I book Thursday?”, raw reasoning power is rarely the bottleneck. HighLevel’s own pitch is to choose “based on speed, cost or reasoning capabilities” — and for front-desk questions, the difference you’re more likely to notice is voice — how long the replies run, how warm they sound, how hard they push toward the booking link, how they handle someone who’s annoyed.

Which is why the right way to pick isn’t a benchmark chart. It’s three of your own real messages, run on two models, in a test sub-account:

• The pricing question. “How much does it cost?” — does the agent answer from your knowledge base, or wander into a generic paragraph?
• The booking request. “Can I come in next week, maybe Tuesday afternoon?” — this also tests the relative-date upgrade HighLevel mentions.
• The angry reply. “I’ve called three times and nobody picked up.” — does it apologise like a person, or like a terms-of-service page?

Run all three on your current model, then switch the dropdown and run them again with the same prompt and knowledge base. Read the replies side by side, out loud if you can, and ask one question: which one sounds like the business, and which one sounds like a chatbot?

Then go back to the tone calibration section of the setup post. A new model can quietly undo tone work you did on the old one — a prompt tuned to rein in a chatty model can make a terse one sound curt. Retest, don’t assume.

The fallback setting nobody will turn on

Picture a dental practice at 11pm on a Tuesday. A patient with a cracked tooth texts the after-hours number. The Conversation AI agent is supposed to answer, triage and offer the first slot tomorrow. Except tonight, the model provider behind that agent is having a bad hour.

Without a fallback, that patient gets silence, or a very slow reply. With one, HighLevel’s own description is that a secondary model will “automatically take over if the primary provider experiences downtime or latency” (Release Radar, Oct 19 week). The patient gets an answer. Maybe a slightly differently-worded one. They don’t care. They got booked.

That’s the whole argument. A slightly different answer beats no answer, every time. And you can only get it by picking a primary and a secondary — the feature is optional, which means most people will leave it blank.

My rule of thumb — reasoning, not a measured result: put the fallback on a different provider than the primary. If your primary is Claude, make the fallback an OpenAI or Gemini model, and vice versa. A fallback from the same provider can be affected by the same outage, which rather defeats the point.

And once you’ve set it, run your three test prompts on the fallback model too. It’s going to be talking to your leads on the worst night of the quarter. You want to know what it sounds like before then.

”Labs” is doing a lot of work in that sentence

You’d think the release notes would end there. They don’t — and the word that matters most is the one in step one. This feature lives in Labs, HighLevel’s opt-in area for newer features. The Release Radar doesn’t call it beta, but it doesn’t call it generally available either, and you turn it on there rather than finding it on by default.

Translation for agencies: don’t write a client SLA around it yet. Don’t tell a client “your bot now runs on Claude” in a proposal. Labs features can change shape, move, or get renamed between the week you test them and the week you’d want to rely on them.

What I’d do instead:

• Turn it on in one sub-account — ideally your own or a demo account — not across every client.
• Switch one agent, keep the others on their current model, and compare a week of real conversations.
• Only then roll it out client by client, with a note in your build log of which agent runs which model. Six months from now, nobody will remember.

If you’re new to how HighLevel’s AI pieces fit together — Conversation AI, Voice AI, the workflow AI actions — the AI workflows pillar maps it out.

What it costs, and why I’m not guessing

HighLevel’s Release Radar post says nothing about pricing for these models. So let’s look at what is published.

The AI product pricing article (checked Oct 2026) lists AI Employee Growth at $50/month and Unlimited at $97/month, both per enabled location, plus a Pay-Per-Use option. On Pay-Per-Use, Conversation AI is billed on token usage — input and output tokens at each model’s per-million-token rate. And here’s the gap: the Conversation AI table in that article still lists only OpenAI models. Anthropic and Google models show up in the Voice AI table, not the Conversation AI one.

So as of today, there’s no published Conversation AI rate for Claude or Gemini that I can point you to. Which means:

• If you’re on Pay-Per-Use, check your AI usage after switching a busy agent, before you switch the rest.
• If you’re on Growth or Unlimited, check whether your included usage behaves differently on the new models.
• If you resell AI to clients, don’t change your client pricing until HighLevel publishes the rates.

I’ll update this section when HighLevel adds Claude and Gemini to the Conversation AI pricing table. If you’re weighing the AI Employee plan against the rest of your GoHighLevel bill, the real cost breakdown puts it in context.

Who should flip the toggle this week, and who should wait

Flip it this week if you already have a Conversation AI agent live, handling real conversations, with a prompt you’re happy with. You have something to compare against. Set a fallback even if you keep your current primary model — that alone is worth the ten minutes.

Wait if you’re still building your first agent. Get the knowledge base, booking flow and tone right on the default model first. Swapping models on an agent that isn’t finished just gives you two unfinished agents. And if you’re still deciding whether a bot should be answering your phones at all, read when AI Employee beats a human receptionist — and the AI Employee review for the wider picture — before you touch any of this.

FAQ

Can GoHighLevel Conversation AI use Claude or Gemini?

Yes, per HighLevel’s Release Radar for the week of October 19, 2026. Turn on “Claude & Gemini Model Support” in Labs, then edit a Conversation AI agent and pick a Claude or Gemini model from the model dropdown. The choice is per agent, so different agents can run on different providers.

Which Claude and Gemini models are available in Conversation AI?

The Release Radar post says seven new models, and names four of them: Claude Opus 5.5, Claude Sonnet 5.5, Gemini 3.8 Flash and Gemini 3.5 Flash Lite. Check the model dropdown in your own account for the full current list, because Labs features can change.

What is the secondary (fallback) model in Conversation AI?

It’s an optional second model you configure on the agent. According to HighLevel, it automatically takes over if the primary provider experiences downtime or latency. It’s the cheapest insurance in the feature: set it to a model from a different provider than your primary.

Does switching to Claude or Gemini cost more in GoHighLevel?

HighLevel hasn’t said. Its AI pricing article (checked Oct 2026) bills Conversation AI on Pay-Per-Use by tokens at each model’s rate, but the Conversation AI table still lists only OpenAI models. Until Claude and Gemini rates appear there, watch your own usage after switching rather than assuming a price.

Is Claude and Gemini support in Conversation AI generally available?

Not as of late October 2026. It’s turned on through Labs, which is HighLevel’s opt-in area for newer features. Treat it as something to test in one sub-account first, not something to build a client commitment on.

What to do this week

Three things, in this order. Turn on the Labs toggle in one test sub-account. Run three of your real messages — a pricing question, a booking request, an angry reply — through your current model and one new one, and compare. Then set a fallback from a different provider, whichever primary you land on.

Let everyone else argue about which model is smartest. The question that matters at 11pm on a Tuesday is simpler: when the primary goes quiet, does anything answer? Make sure yours does.