AI Agent Pricing in 2026: 40 Tools Analyzed
We researched the pricing model, free tier and API access for all 40 AI agents in our index — freemium leads at 40%, and 80% offer some free access.
By The Agents Index Editorial, Research desk · 2026-07-19
We pulled the researched pricing model, free-tier status, API availability, open-source status, interface and deployment for every one of the 40 Published listings in our own index — the same structured facts each listing page carries, gathered in one place. Snapshot date: 21 July 2026. No single pricing model dominates the category the way per-seat SaaS once did: freemium is the largest single group at 43%, but usage-based billing (28%), permanently free tools (18%) and flat subscriptions (13%) are all real, active patterns — and which one a given tool uses tracks its category far more than its size or funding.
Intraday correction (21 July 2026): Cline moved from "free (no paid tier)" to "freemium" the same day as this snapshot, after live re-verification found it now sells an optional $9.99/month ClinePass subscription and a custom-priced Enterprise tier (SSO, SLA, dedicated support, the JetBrains plugin) alongside its free, open-source core. Every count and percentage below reflects that correction — this is not a new quarterly edition, just this report staying in sync with its own source listing on the day the underlying fact changed.
What changed since the last snapshot (19 → 21 July 2026): the index grew from 30 to 40 Published listings (+33%), crossing this report's own stated 25%-growth refresh trigger. All 10 additions now carry a full pricing datasheet — Elicit, ElevenLabs Conversational AI, Juggler, OpenAI Codex CLI, OpenCode, Zot, Artisan, Decagon, Crescendo and Qualified — and OpenHands, which launched through the submission pipeline without one, has since had its datasheet backfilled. That means this edition is the first with a genuinely uniform N=40 dataset and no asterisked exception. The headline shape barely moved (freemium 41%→40%), but two real shifts showed up: free access dropped from 87% to 80% as four of the ten new tools (all enterprise-sales-led) charge from the first use, and voice/support agents' 100%-usage-based pattern held on a fuller sample (now 4/4 each, was 3/3 and 2/2).
Key findings
- Of the 40 AI agents in our index, 17 (43%) use freemium pricing — the single largest group, but not a majority.
- 11 of 40 (28%) bill by usage or outcome rather than a flat fee — the second-largest group, and a genuinely common model, not an edge case.
- 7 of 40 (18%) are free with no paid tier at all, and 5 of 40 (13%) are a flat subscription with no usage-based component.
- 32 of 40 tools (80%) offer some form of free access before paying — a permanently free tool, a freemium tier, or a capped free trial. Eight tools require payment from the first use — double the four in the last snapshot: 11x, Claude Code, Lindy and Sierra as before, joined by four new enterprise-sales-led additions — Artisan, Decagon, Crescendo and Qualified.
- Usage-based pricing is a category signature, not noise — and it just got more unanimous. All 4 voice agents (Bland AI, Retell AI, Vapi, ElevenLabs Conversational AI) and all 4 support agents (Intercom Fin, Sierra, Decagon, Crescendo) in our index bill by usage or outcome; zero of either group is a flat subscription.
- Conversely, every one of the 6 agent frameworks in our index (LangChain, CrewAI, Microsoft AutoGen, LlamaIndex, OpenAI Agents SDK, Pydantic AI) is free or freemium — none charges a flat subscription or usage fee for the framework itself (paid observability/hosting add-ons exist, e.g. LangSmith, but the core is always free).
- Among the 40 tools, the average is 3.1 published pricing tiers, and 24 of 40 (60%) offer a named Enterprise or custom-priced tier.
- 25 of 40 tools (63%) expose a public API — the live export and the research datasheet now agree exactly on every one of the 40 rows, so this report drops the two-granularity caveat earlier editions needed.
- 16 of 40 tools (40%) are open source; self-hosting concentrates almost entirely inside that set — every self-host-only or local-only tool in our index is also open source.
- 16 of 40 tools (40%) are model-agnostic — you bring your own LLM API key (Claude, GPT, Gemini, a local model, or otherwise) rather than being locked to one vendor's model.
How the market splits: no single model wins
All 40 Published listings in the index now carry a full pricing datasheet, so this table is a complete census, not a partial sample:
| Pricing model | Tools | Share of 40 |
|---|---|---|
| Freemium | 17 | 43% |
| Usage-based | 11 | 28% |
| Free (no paid tier) | 7 | 18% |
| Flat subscription | 5 | 13% |
Freemium is the largest single bucket, but at 43% it is nowhere near a majority — a buyer researching "AI agent pricing" should expect to run into a genuinely usage-billed or flat-subscription product about as often as a classic try-then-upgrade freemium one.
Free access is the norm — but the gap is narrowing
32 of 40 tools (80%) let a buyer use something for nothing before paying — either a permanently free tool (7 tools: Aider, GPT Researcher, Microsoft AutoGen, OpenAI Agents SDK, Pydantic AI, Juggler and Zot) or a freemium/usage-based tier with a genuine free starting point (25 tools, including Cline, which added an optional paid ClinePass/Enterprise layer alongside its free core in mid-2026). Eight tools in the index require payment before a buyer can use anything, twice as many as the last snapshot: 11x (an AI SDR platform sold on custom enterprise contracts), Claude Code (bundled into a paid Claude Pro/Max subscription or billed per API token), Lindy (a no-code agent-building platform on paid plans only), Sierra (an enterprise customer-service platform sold on outcome-based contracts), Artisan (an AI SDR platform, subscription-only), Decagon and Crescendo (both enterprise support agents, sales-led with no self-serve tier), and Qualified (a Salesforce-native inbound AI SDR, now Salesforce-owned, custom contracts only). All four new no-free-access tools share the same shape: enterprise support or sales agents sold through a sales demo rather than a self-serve signup — the entire growth in this group came from that one buying motion, not from the index broadening at random.
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? 25 of 40 (63%) — the live public export and the fuller research datasheet agree on every single row now that OpenHands' datasheet gap has been backfilled, so this edition needed only one figure instead of two.
- Is the source code open? 16 of 40 (40%) are open source. Every framework in our index (LangChain, CrewAI, Microsoft AutoGen, LlamaIndex, OpenAI Agents SDK, Pydantic AI) is open source, as are all four of the newest CLI-native coding agents (Juggler, OpenAI Codex CLI, OpenCode, Zot); most hosted platform and voice-agent products are not.
- Can you bring your own model? 16 of 40 (40%) are explicitly model-agnostic — the tool's own documentation states you supply a Claude, GPT, Gemini, local, or other LLM API key rather than being locked into one vendor's model. This overlaps with but is not identical to open source: OpenAI Agents SDK and OpenAI Codex CLI are both open source but OpenAI-first, not counted as agnostic; Aider, OpenCode and Zot are open source and fully model-agnostic.
Deployment now shows a genuinely new shape: 22 of 40 (55%) are cloud-only, 11 of 40 (28%) are self-host-only or local-machine-only, and 7 of 40 (18%) support a real hybrid. The "local-only" and "local + cloud" values are new to this edition entirely — three CLI-native coding agents (OpenCode and Zot run local-only; OpenAI Codex CLI runs local or cloud) weren't a pattern the index had in the last snapshot. Self-hosting and local-only still concentrate almost entirely in the open-source set — every self-host-only or local-only tool in our index is also open source.
What you build predicts how you're priced
The clearest pattern in the dataset is that pricing model tracks category, not company size or funding. We cross-referenced each tool's category against its pricing model:
| Category | Members | Dominant pricing model |
|---|---|---|
| Coding agents | 14 | Freemium (10/14, 71%) |
| Frameworks | 6 | Freemium or free (6/6, 100% — zero paid) |
| Agent platforms | 5 | Freemium (2/5, 40%; the rest split usage-based/subscription) |
| Research agents | 4 | Freemium (2/4, 50%; usage-based and free split the rest) |
| Sales & marketing agents | 4 | Subscription (3/4, 75%) |
| Voice agents | 4 | Usage-based (4/4, 100%) |
| Support agents | 4 | Usage-based (4/4, 100%) |
The split reads as a straight line from what a category of agent actually does. Frameworks are developer building blocks — you already pay for the compute and the model API, so the framework itself competes on adoption, not licensing, and every single one in our index is free or freemium. Voice and support agents sell a completed outcome (a resolved call, a resolved ticket) with a variable, real per-unit cost behind it — usage- or outcome-based pricing is the only model that doesn't force the vendor to either under-price heavy users or over-price light ones. Sales & marketing agents flipped the other way this edition: with two more enterprise AI-SDR tools added (Artisan, Qualified), flat subscription is now the category's dominant model (75%) instead of an even split — the category is trending toward sold-not-self-served as it fills out. Coding agents still sit in the middle: freemium remains dominant (71%) because the product benefits from a large free user base for distribution, but the four newest coding-agent additions are split between fully free CLI tools (Juggler, Zot) and freemium CLI tools (OpenAI Codex CLI, OpenCode) — the CLI wave hasn't picked a single pricing shape yet.
The full dataset
Every row below is pulled directly from the same live, gate-passed listing data our individual tool pages render from — nothing here is estimated. All 40 rows now carry a complete structured datasheet, so — unlike the last edition — no row needs a "—" placeholder.
| Agent | Category | Pricing model | Free tier | Public API | Open source | Deployment |
|---|---|---|---|---|---|---|
| 11x | Sales/Mktg | Subscription | No | Yes | No | Cloud |
| Aider | Coding | Free | Yes | No | Yes | Self-host |
| Artisan | Sales/Mktg | Subscription | No | No | No | Cloud |
| Bland AI | Voice | Usage-based | Yes | Yes | No | Both |
| Bolt.new | Coding | Freemium | Yes | No | No | Cloud |
| Claude Code | Coding | Subscription | No | Yes | No | Cloud |
| Clay | Sales/Mktg | Usage-based | Yes | Yes | No | Cloud |
| Cline | Coding | Freemium | Yes | Yes | Yes | Self-host |
| Crescendo | Support | Usage-based | No | No | No | Cloud |
| CrewAI | Frameworks | Freemium | Yes | Yes | Yes | Self-host |
| Cursor | Coding | Freemium | Yes | No | No | Cloud |
| Decagon | Support | Usage-based | No | No | No | Cloud |
| Devin | Coding | Freemium | Yes | Yes | No | Cloud |
| Dify | Platforms | Freemium | Yes | Yes | Yes | Both |
| ElevenLabs Conversational AI | Voice | Usage-based | Yes | Yes | No | Cloud |
| Elicit | Research | Freemium | Yes | Yes | No | Cloud |
| GitHub Copilot | Coding | Freemium | Yes | No | No | Cloud |
| GPT Researcher | Research | Free | Yes | Yes | Yes | Self-host |
| Intercom Fin | Support | Usage-based | Yes | Yes | No | Cloud |
| Juggler | Coding | Free | Yes | No | Yes | Self-host |
| LangChain | Frameworks | Freemium | Yes | Yes | Yes | Self-host |
| Lindy | Platforms | Subscription | No | No | No | Cloud |
| LlamaIndex | Frameworks | Freemium | Yes | Yes | Yes | Both |
| Manus | Research | Usage-based | Yes | Yes | No | Cloud |
| Microsoft AutoGen | Frameworks | Free | Yes | No | Yes | Self-host |
| n8n | Platforms | Freemium | Yes | Yes | Yes | Both |
| OpenAI Agents SDK | Frameworks | Free | Yes | No | Yes | Self-host |
| OpenAI Codex CLI | Coding | Freemium | Yes | Yes | Yes | Local + Cloud |
| OpenCode | Coding | Freemium | Yes | Yes | Yes | Local |
| OpenHands | Coding | Freemium | Yes | Yes | Yes | Both |
| Perplexity | Research | Freemium | Yes | Yes | No | Cloud |
| Pydantic AI | Frameworks | Free | Yes | No | Yes | Self-host |
| Qualified | Sales/Mktg | Subscription | No | Yes | No | Cloud |
| Relevance AI | Platforms | Usage-based | Yes | Yes | No | Cloud |
| Replit Agent | Coding | Freemium | Yes | No | No | Cloud |
| Retell AI | Voice | Usage-based | Yes | Yes | No | Cloud |
| Sierra | Support, Platforms | Usage-based | No | No | No | Cloud |
| Vapi | Voice | Usage-based | Yes | Yes | No | Cloud |
| Windsurf (now Devin Desktop) | Coding | Freemium | Yes | No | No | Both |
| Zot | Coding | Free | Yes | Yes | Yes | Local |
Methodology
Dataset: all 40 Published listings in The Agents Index as of 21 July 2026, read live from our own public export (theagentsindex.com/listings.json) and cross-checked against the underlying research datasheet for all 40 — the first edition where every row carries a complete datasheet, since OpenHands' gap (the one exception in the last snapshot) has since been backfilled.
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 (commercial add-ons like LangSmith or Logfire don't count against it); "model-agnostic" means the tool's own documentation states you can supply your own LLM API key rather than being locked to one vendor's model. Percentages are rounded to the nearest whole number, so figures in a table may not sum to exactly 100.
Limitations: this is a census of our own index, not the entire AI-agent market — a 40-tool, 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 report reflects a snapshot, not a live feed. Refresh cadence: quarterly, or sooner if the listing count grows more than 25% since the last publish (this edition: 30→40, +33%, which is what triggered it).
Common questions
What percentage of AI agents are free to try?
Of the 40 AI agent tools in our index, 32 (80%) offer some form of free access — either a permanently free tool with no paid tier, a freemium plan with a free starting tier, or a capped free trial. 8 tools require payment before you can use anything, twice as many as our last snapshot.
Are most AI agent tools open source?
No — 16 of the 40 tools in our index (40%) are open source. Openness concentrates heavily in two places: every agent framework (LangChain, CrewAI, Microsoft AutoGen, LlamaIndex, OpenAI Agents SDK, Pydantic AI) and every CLI-native coding agent (Juggler, OpenAI Codex CLI, OpenCode, Zot) is open source, while most hosted platforms and voice agents are closed source.
Do AI agent tools usually charge for API access?
Slightly more than half do. 25 of 40 tools in our index (63%) expose a public API; the remaining 37% are closed products with no documented API for third-party integration.
Which kind of AI agent is most likely to charge by usage instead of a flat fee?
Voice agents and support agents. Every voice agent (4/4) and every support agent (4/4) in our index bills by usage or resolved outcome rather than a flat subscription — the opposite of agent frameworks, where every tool (6/6) is free or freemium with no usage-based tier at all.
Every listing referenced above carries its own sourced pricing tiers, verification date and editorial verdict — start with the full ranked index or browse by category. For how we keep this data honest in the first place, see how we research AI agents and our methodology.