AI Agent Pricing in 2026: Every Tool in Our Index

We pulled the researched pricing model, free-tier status, API availability, open-source status and deployment for every one of the 92 Published listings in our own index across coding, framework, platform, voice, research, sales-and-marketing and support agents: the same structured facts each listing page carries, gathered in one place. Snapshot date: 27 August 2026. No single pricing model dominates the category the way per-seat SaaS once did: freemium is the largest single group at 46% (42 of 92), but usage-based billing (26%, 24 tools), flat subscriptions (17%, 16 tools) and permanently free tools (11%, 10 tools) are all real, active patterns. Which one a given tool uses tracks its category far more than its size or funding.

All 92 rows appear in the table below, and every aggregate figure in this report is recomputed from that same live set in a single pass, so no numerator is ever left standing against a denominator it was not counted with. Some fields are genuinely unresearched rather than assumed negative; see the Openness section below for what that affects.

Key findings

• Of the 92 AI agents in our index, 42 (46%) use freemium pricing: the single largest group.
• 24 of 92 (26%) bill by usage or outcome rather than a flat fee. That is the second-largest group, and a genuinely common model, not an edge case.
• 16 of 92 (17%) are a flat subscription with no usage-based component, and 10 of 92 (11%) are free with no paid tier at all. Those are the smallest groups.
• 57 of 92 tools (62%) offer some form of free access before paying: a permanently free tool, a freemium tier, or a usage-based tier with a genuine, ongoing free starting point.
• 47 of 90 tools (52%) were model-agnostic when we classified them on 19 August 2026. You bring your own LLM API key, or an admin can select and switch between multiple hosted providers, rather than being locked to one vendor's model. That one figure is a dated reading rather than a live count: our corpus records each tool's model stack as prose, and the structured field that will make it live was only added on 20 August 2026, so only 1 of 92 carry a value so far. The figure goes live here once all 92 do.
• 69 of 92 tools (75%) are confirmed to expose a public API, and 10 are confirmed not to. The remaining 13 carry no recorded value for this attribute, so they are counted separately rather than folded into "No."

How the market splits: no single model wins

Free access is common, but a real minority pays from the first click

57 of 92 (62%) let you use something without paying: a permanently free tool, a freemium tier, or an ongoing free allowance under usage-based billing. The other 35 (38%) want a card, a contract or a sales call before you see the product work on your own data.

Openness: APIs, open source, and bring-your-own-model

Three separate questions get conflated under "openness," and our data keeps them apart.

Does it expose a public API? 69 of 92 (75%) are confirmed "Yes" and 10 are confirmed "No." The remaining 13 have no recorded value: unresearched, not negative.

Is it open source? 33 of the 91 tools with a recorded value (36%) ship their core agent code under an OSI-style license.

Can you bring your own model? 47 of 90 tools (52%) were model-agnostic when we classified them on 19 August 2026. You bring your own LLM API key, or an admin can select and switch between multiple hosted providers, rather than being locked to one vendor's model. That one figure is a dated reading rather than a live count: our corpus records each tool's model stack as prose, and the structured field that will make it live was only added on 20 August 2026, so only 1 of 92 carry a value so far. The figure goes live here once all 92 do.

Deployment: 51 of 92 (55%) are cloud-only, 13 of 92 (14%) are self-host-only or local-machine-only, and 28 of 92 (30%) support a real hybrid. In total, 41 of 92 (45%) have some form of self-host path.

What you build predicts how you're priced

Support agents are the single most consistent category on pricing model: every one of the 8 bills by usage or resolved outcome, not a flat seat fee. Voice agents follow the same logic at 75%: you pay per minute of a call, the same way you'd pay per resolution for a support ticket. Frameworks sit at the other end: 71% are freemium, because the cost that matters to a framework's buyer is the model API bill underneath it, not the SDK itself. Sales and marketing agents split the other way from voice and support: five of the eight (11x, Amplemarket, Artisan, Qualified and Warmly) are sales-led flat subscriptions rather than metered by call or resolution, because the buyer is provisioning a rep's worth of pipeline work rather than paying per contact touched. The freemium minority in the category is the self-serve set: Rox, Unify and Lead Scorer. Each ships a genuine $0 or €0 tier a solo operator can run without a sales call, with Lead Scorer's free CRM and MCP access sitting alongside a Pro tier that starts at €99/month (about $111) once the autonomous SDR itself turns on.

The full dataset

Every row below is pulled directly from the same live, quality-gated listing data our individual tool pages render from, and nothing here is estimated. All 92 rows carry a structured datasheet, except where marked "Unknown": that means the attribute has no recorded value in our corpus yet, not that the answer is "No."

Methodology

Dataset: all 92 Published listings in The Agents Index across the coding-agents, frameworks, agent-platforms, voice-agents, research-agents, sales-marketing-agents and support-agents categories, as of 27 August 2026, read live from our own tenant database. Every aggregate figure on this page is recomputed from that full set in a single pass, with numerator and denominator counted together, and the page is published verbatim from what it renders. No figure here is hand-typed.

Definitions: "freemium" means a usable free tier plus at least one paid tier; "free" means no paid tier exists at all; "usage-based" means billing scales with consumption or outcomes rather than a flat recurring fee; "open source" means the core agent code ships under an OSI-style license. Percentages are rounded to the nearest whole number, so a column may not total exactly 100.

One figure is dated rather than live: model-agnostic. Our corpus recorded each tool's model stack as prose rather than as a structured flag, and the distinction that matters to a buyer (can you supply your own key, or is this a managed stack that merely happens to route across several models) is a judgement about the vendor's own wording. Classifying it mechanically scores differently from reading it, so the published figure is carried with the date it was made instead of being recomputed into a number nobody researched. A structured modelAgnostic field was added on 20 August 2026 and fills on each re-verify; this paragraph disappears and the figure becomes a live count as soon as any listing carries a value.

Limitations: this is a census of our own index, not the entire AI-agent market, and it is a quality-gated, individually-researched set, not a scrape. A tool's pricing can change after our last verification date (shown on its own listing page). This page is regenerated whenever the corpus moves, so its coverage and its figures cannot drift apart.