Our own GHL receipts: $0.142, then $0.215, then $0.3497.
Three real invoices. Same GoHighLevel agency account. Same agent, with configuration iterated between calls and a model change between months. The per-minute cost climbed roughly two-and-a-half times across the set, and the only thing that changed was what we asked the agent to do.
Most voice-AI pricing debates are theoretical. This one isn't. We pulled three actual call records off our own GoHighLevel agency billing ledger and laid them next to each other. Same account, same OAuth token, same conversation inbox. What follows are the numbers and the lesson under them.
Receipt 1: $0.142/min — the light agent
57-second inbound call. 27,249 input tokens across the LLM calls fired during that minute. The agent was light — a small prompt, a short knowledge base, basic qualifier flow. Effective per-minute rate: $0.1421.
On its own this looks fine. We can sell the client $0.30/min with a comfortable margin, the math works, the demo demo'd. If every minute came back at $0.142, this entire post wouldn't exist.
Receipt 2: $0.215/min — seven minutes later
The same agent on the same account, seven minutes later, made another inbound call. 59-second duration. Input tokens for the minute: 69,066. Effective per-minute rate: $0.2149.
Same agent. The configuration was iterated between calls — we added two knowledge-base entries and tightened a couple of objection-handling patterns. The token footprint went from 27K to 69K because the agent now had more context to reach into. The platform charged us for that linearly. The per-minute cost went up 51% in seven minutes, and there was no warning in the dashboard that it would.
This is the part that surprises agencies. The platform isn't doing anything wrong — they're charging by tokens, and we added tokens. But the agency-economics implication is that configuring the agent is a billable event, indefinitely, every minute of every call from now on. There's no "set the agent once and the rate is the rate" mode.
Receipt 3: $0.3497/min — two months earlier, on GPT-4.0
This one's the original receipt. March 25, 2026. A 10:44 call on the same agency account when we were still running the agent on GPT-4.0. Invoice total: $3.7535 for one call. Effective per-minute rate: $0.3497.
What's interesting about this one isn't just the number — it's that the rate landed inside GoHighLevel's own pricing estimator. We went back into the GHL builder, reverted the agent to GPT-4.0, and the estimator quoted us $0.306 – $0.420/min for that configuration. The estimator was honest. The actual call landed at $0.3497, comfortably inside the quoted band.
That's the part you want to sit with. The platform didn't lie. The rate was visible up front. It still wasn't survivable for a flat-priced client retainer.
Three lessons that compound
1. Model swaps reduce the rate. They don't fix the meter.
We moved off GPT-4.0 to gpt-5.2 (and Haiku is now an option on GHL too). The per-minute cost on the same agent dropped from $0.35-ish to $0.14–$0.21. That's a real savings. But the structure didn't change — the bill still moves at call time based on tokens consumed during the call. Two months later we measured a 51% swing on the same Haiku-era agent within seven minutes.
2. The agent's capability is the meter's input variable.
Adding a knowledge-base entry, adding a tool, tightening a script — every one of those grows tokens. The richer the agent, the higher the per-minute floor. This is the opposite of how you want infrastructure to behave when you're selling outcomes to agencies.
3. The caller has more leverage on your bill than you do.
You can decide what tokens to bake into the agent at build time. You cannot decide whether the next caller is going to talk for forty seconds or four minutes, whether they're going to trigger the knowledge base twelve times or twice, whether they're a robocaller racking up minutes against a misconfigured forwarding rule. Your downside is unbounded; the caller controls the upper end.
Why we built Callibre this way
Callibre wasn't designed to compete on price. It was designed to make the per-minute cost stop moving at call time. That's the entire shape of the product. Voice AI is a flat $0.07/min. LLM is a flat per-minute line you pick at build time. Telephony is pass-through at $0.010/min. Add them up and you get a single per-minute number — $0.086 to $0.140 depending on the LLM — that's locked the moment you ship the agent.
What that means in agency terms: the same call that drifted from $0.142 to $0.215 over seven minutes on our GHL receipts would have been a single Callibre line at $0.105/min (Haiku default), unchanged across both invoices. The bill stops being a variable.
The three receipts above are on the comparison page with the actual conversationAI reference IDs, timestamps, and the GHL pricing-estimator screenshot that confirmed the quoted band. If you want to drop your client's call volume into the calculator and see the spread, it's on the same page.
Where this leaves us
None of this is a takedown of GoHighLevel. We built Callibre inside the HighLevel ecosystem, on top of HighLevel integrations, in partnership with HighLevel-native agencies. The metered voice-AI pricing model is rational from the platform's side. It just isn't the right shape for an agency that's quoting clients a flat per-minute rate and trying to defend a margin floor twelve months from now.
Three receipts. One agency. One structural lesson. Pick the platform that makes your math survive contact with month three.
See the receipts side by side.
Three GHL invoices, the GPT-4.0 estimator exhibit, and the Callibre line at $0.105/min for the same calls.
Read the full comparison →