Per-Message Pricing: Examples & Companies

What is it

Per-message pricing is a billing unit where each individual message or reply in a conversation is metered, common in AI chat and voice platforms.

The message is the most intuitive unit in conversational AI. It mirrors how users already think — “I sent a message, I got a reply” — and maps onto the single LLM call that drives inference cost. Where token pricing is precise but opaque to non-technical buyers, message pricing trades granularity for a meter that is human-readable before the buyer opens the docs.

The unit shows up in two commercial contexts. Consumer AI platforms — Poe, Grok, Janitor AI, and the now-sunset Cognosys — use message caps to differentiate subscription tiers rather than billing per message, so the message rate is the invisible floor even when no dollar-per-message line item ever appears. Voice and chat API platforms take the opposite tack: Retell AI and Vapi publish explicit per-message rates for their chat-agent products — $0.002 and $0.005 respectively — treating the message as a pure usage meter alongside per-minute voice billing.

Edge cases reveal how elastic “message” can be. Intercom Fin carries messages through a Proactive Support Plus add-on, separate from its headline per-resolution model, and Pi rate-limits on messages purely to control cost — its effective rate is $0, and the meter never converts to a charge. A message meter in the frontmatter does not imply a per-message bill. For how to pick a unit like this, see choosing the right usage metric.

How it works

A message meter counts conversational turns and multiplies by a rate. The rate can be an explicit dollar figure (API platforms), a points cost that varies by model (Poe), or an implicit ceiling encoded as a tier cap (consumer apps). What differs across vendors is what actually gets counted and how the price is exposed.

The clearest direct-metered example is Retell AI. Its chat agents are billed per AI message from $0.002, and the effective rate depends on the model — a GPT 5.1 chat message computes to about $0.013. There is no platform fee, no contract, and new accounts get $10 in free credits.

Unit math (Retell chat agent): 5,000 GPT 5.1 messages × $0.013/msg ≈ $65. The same volume at the $0.002 floor rate (a lighter model) is $10 — the model choice is the dominant cost lever, not the message count.

Poe shows the points-abstraction variant: one subscription buys a monthly (or daily) compute-points allowance, and every message spends points at a published, model-dependent rate. A lightweight text model costs ~10–20 points per message, so a $19.99 plan stretches to tens of thousands of cheap messages — but a single high-end video generation can cost more points than thousands of text messages. The message is the meter; the model mix is the real cost driver.

Companies using this

Eight companies in the corpus list messages as a billing unit, split between voice/chat API platforms that meter it directly (Retell AI, Vapi) and consumer apps that use message caps to gate subscription tiers (Poe, Grok, Janitor AI, Cognosys).

Patterns observed

• Direct metering lives in the API layer, caps live in the app layer. The only companies quoting a public per-message dollar rate — Retell AI and Vapi — are developer platforms selling chat agents. Every consumer product expresses the unit as a tier cap instead: Cognosys at 100 messages/month, Janitor AI at ~50/day, Grok at ~10 prompts per 2 hours.

• The message is almost never the only meter. On every company here, messages rides alongside another unit — minutes for Retell AI and Vapi, points for Poe, tokens for Grok and Janitor AI, resolutions for Intercom Fin. Message pricing is a legibility layer over an underlying token or compute cost, not a standalone model.

• Model choice, not message volume, drives the bill. Where the rate is model-dependent — Poe’s points, Retell AI’s per-model chat rates — the same message count can cost 5–10× on the model selected. Buyers budgeting on message count alone consistently underestimate frontier-model spend.

Counterexamples & variants

The clearest counterexample is Pi. It lists messages as a billing unit, but the app has never been monetized — Inflection AI rations capacity with rate limits and cooldowns rather than dollars, so the effective rate is $0 and the meter never converts to a bill. It is a pure cost-control meter, not a pricing model.

Cognosys is a lifecycle variant: it ran a clean message-metered freemium ladder (Pro ~1,000 msg/mo at $15, Ultimate unlimited at $59) but is now sunset after the Cohere acquisition, so its rate card is historical rather than purchasable. Intercom Fin inverts the emphasis: its economics run on $0.99-per-resolution outcome pricing, with messages surfacing only in a peripheral Proactive Support add-on (500/mo at $99). The message unit is real but commercially minor.

What this means for buyers vs vendors

For buyers

Confirm exactly what a “message” is before you model spend — Retell AI counts each AI reply, Poe charges per message to a bot, and consumer caps count prompts. Then model on your worst-case model mix, not message volume, since a model-dependent rate is what actually moves the bill. Use the pricing calculator to stress-test a high-volume month and read usage invoicing and billing cycles to understand how overages and caps are reconciled.

For vendors

Per-message billing fits when your unit cost is predictable per exchange and your buyer is non-technical enough that tokens would confuse them — the message is a legibility win that Retell AI turned into a self-serve growth lever. Pair it with a second meter that tracks real cost (minutes, points, or tokens) so a frontier-model exchange doesn’t erode margin, and start with the fundamentals in introduction to usage-based pricing. Expect to publish caps or free-tier limits, since every consumer app on this page gates messages rather than charging for them directly.

Explore this theme in the knowledge graph →