AI Agent Pricing Models: 24 of 89 Tools Bill by Usage

Of the 89 AI agents published in our index on 18 August 2026, 24 bill by usage, 38 are freemium, 17 are flat subscriptions and 10 are free. The classification records what each vendor’s own pricing page uses as its headline meter, which is the unit the bill scales with.

The distribution is not the useful part. The useful part is that the usage meter is not spread across the index at all: it sits almost entirely in the agents that talk to your customers, and almost nowhere in the agents that sit beside your developers.

The headline count, out of 89

Every listing in our index carries a Pricing Model field set to one of four values: free, freemium, subscription, or usage-based. On 18 August 2026 the field was populated for 89 of 89 published listings, so every count below is out of the whole index and there is no unknown bucket to set aside.

• 38 of 89 (43%) are freemium, with a real free tier or a free open-source core beneath the paid tiers
• 24 of 89 (27%) are usage-based, where the headline meter is a unit of work: a minute, a resolution, an action, a credit
• 17 of 89 (19%) are subscription, a flat recurring price per seat or per plan with no free tier
• 10 of 89 (11%) are free, with no paid tier of any kind

The meter follows who the agent talks to

Splitting the index by category shows where each meter lives. The row totals sum to 91 rather than 89 because two listings, Sierra and Botsify, sit in two categories each and are counted once in both. Sierra is usage-based and Botsify is a subscription, which is also why the usage-based column sums to 25 and the subscription column to 18.

Category	Freemium	Usage-based	Subscription	Free	Total
Coding agents	14	2	4	4	24
Agent frameworks	10	0	0	4	14
Agent platforms	5	1	5	1	12
Voice agents	2	8	1	0	11
Research agents	4	5	1	1	11
Support agents	0	8	1	0	9
Sales & marketing agents	2	1	5	0	8
Agent tools	1	0	1	0	2

Voice agents and support agents hold 16 of the 25 usage-based category slots between them. Frameworks hold none at all.

Customer-facing agents meter, developer-facing agents do not

Take the categories where the agent handles an end customer, meaning voice agents and support agents. No tool sits in both, so that is 20 distinct listings, and 16 of those 20 bill by usage.

All eight usage-based support agents meter a resolved conversation or a handled ticket: Sierra, Intercom Fin, Decagon, Crescendo, Cresta, Ada, Zendesk AI Agents and Auto-Respond. Fin is billed per resolution on Intercom’s own pricing page. All eight usage-based voice agents meter a minute of call time: Vapi, Retell AI, Bland AI, PolyAI, Synthflow AI, ElevenLabs Conversational AI, Deepgram Voice Agent API and the OpenAI Realtime API.

Now take the categories where the agent works alongside an employee, meaning coding agents and frameworks. That is 38 distinct listings with no overlap, and 2 of those 38 bill by usage: Devin and Muse Code. Instead 32 of the 38 are free or freemium, and the remaining four are flat subscriptions: Claude Code, Factory, Augment Code and Gemini Code Assist.

16 of 20 against 2 of 38 is not a gradient. It is the most useful thing on this page for anyone building a budget, because the two shapes fail in opposite directions. A voice or support agent is priced like infrastructure, so the bill scales with how much work it does, which makes it near-free in a pilot and hard to forecast at volume. A coding agent or framework is priced like a tool: free to start, then a flat rate per developer, largely indifferent to how hard the team leans on it.

The reason is visible in what each category can meter. A resolved support ticket and a minute of phone call are units the buyer already counts and already has a human cost for, so a price per unit can be compared against something. There is no equivalent unit for “wrote some code”, because the industry has no agreed meter for developer output, and the seat survives as the fallback.

Open source and metered billing do not mix

Crossing the pricing field against the open-source field gives the cleanest separation in this index, and as of 18 August 2026 it no longer needs a caveat. Open-source status is now determined for 89 of 89 published listings, so this table covers the whole index and there is no undetermined column to set aside.

	Open source	Closed source	Total
Free	10	0	10
Freemium	20	18	38
Subscription	0	17	17
Usage-based	0	24	24
Total	30	59	89

All 30 open-source AI agents in our index are free (10) or freemium (20). Not one of them bills by subscription or by usage. In the other direction, all 10 free tools are open source, and no closed-source tool in the index charges nothing at all.

Half of that is close to a tautology and half of it is genuinely informative. A tool whose source you can run yourself struggles to enforce a meter, so the empty cells in the open-source column are what you would expect. The informative half is the 18 closed-source freemium tools, which are the only group giving something away without being obliged to. The largest cluster of them, 8 of the 18, are coding agents, where a free tier is how a developer tool gets tried at all.

Two earlier versions of this page stated that zero as a claim about 83 listings rather than 89, because six listings had no determined open-source status and four of those six sat in the metered or recurring buckets. Those six were researched on 18 August 2026 and all six came back closed source: Augment Code, Gemini Code Assist and SalesTouch are subscriptions, TinyFish is usage-based, and Browserbase and Julius AI are freemium. The claim survived the check that could have broken it, which is why it is now stated across the whole index rather than a subset of it.

What would falsify it is one open-source vendor putting a metered or flat-rate price on the hosted product with no free path, and that is a real possibility rather than a rhetorical one. This is a count of what 89 vendors publish today, not a law about how open source can be sold.

The 10 free tools have no paid tier of any kind today: Aider, Juggler, Zot, AgentsKit, Goose, Microsoft AutoGen, the OpenAI Agents SDK, Pydantic AI, OpenClaw and GPT Researcher. That describes each vendor’s current published pricing and predicts nothing. Several tools now in the freemium column started here.

A free tier is not the same as freemium

The Free Tier field is also populated for 89 of 89, and its answer is different from the pricing-model field. 54 of the 89 listings offer a free tier of some kind, while only 48 are classified free or freemium.

The gap is six listings that meter or charge a flat fee and still let you start without paying. Five are usage-based: Clay, Manus, Genspark, TinyFish and ElevenLabs Conversational AI. One is a subscription, Augment Code. If the question you care about is whether you can try a tool without a card, the answer is 54, not 48. These are two different questions and their totals do not match.

Method, and two ways to misread this

We read the Pricing Model field on all 89 published listings in this index on 18 August 2026, against each vendor’s own live pricing page. The field was populated for 89 of 89, so no count above is a percentage of a subset. Open-source status was also determined for 89 of 89, the last six having been researched on 18 August 2026, so the open-source table is a whole-index count and not a subset either.

The field records the headline meter: the unit the bill scales with on the vendor’s primary paid path. Where a vendor publishes several tiers, the classification follows what those tiers are denominated in, not the cheapest entry point.

Usage-based does not mean expensive, and subscription does not mean predictable. The field describes the shape of the meter and says nothing about the size of the bill. A per-resolution price can be cheaper than a seat for a small team and far more expensive at volume, which is the point of the distinction rather than an objection to it. Our pricing FAQ works several of these tools through to actual figures, and the entry-price census counts who publishes a number at all.

These counts are a census of this index, not a survey of the market. They describe the 89 tools we have researched and published as of 18 August 2026, the index grows, and the date is part of the claim. Corrections are welcome through our methodology page.