# AI Agent Tools With Usage-Based Pricing

> Source: https://theagentsindex.com/collections/ai-agent-tools-with-usage-based-pricing (curated, quality-gated, re-verified)
> Curated by: The Agents Index Editorial, Research desk

"Usage-based" sounds like the transparent option — pay only for what you use, and calculate your own cost before you sign up. That holds for all but five of these 22 tools, which publish an actual number you can price your own usage against: [Vapi](/vapi) at $0.05/minute, [Bland AI](/bland-ai) at $0.14/minute, [Intercom Fin](/intercom-fin) at $0.99/resolution, [OpenAI Realtime API](/openai-realtime-api) at $32/$64 per million audio tokens, [Webhound](/webhound) from $1 for a narrow research check — a real rate card in every case, whatever the unit. [Devin](/devin) sells the meter as an allowance rather than a rate, $500/month for 250 Agent Compute Units on its Team plan or from $20/month pay-as-you-go, which is still a number you can budget against.

It breaks for the other five. [Sierra](/sierra), [Decagon](/decagon), [Ada](/ada) and [PolyAI](/polyai) each state their pricing model in the abstract — "outcome-based," "per minute," "per conversation" — then list "Custom" where the actual number should be; none publishes a rate card or offers self-serve signup, and independent purchase-data estimates (not vendor figures) are the only cost signal available for three of the four. [Zendesk AI Agents](/zendesk-ai-agents) is the more interesting case: it publishes real seat pricing ($55-$115/agent/month) and bills genuine usage on top through Verified Resolutions, but the overage rate itself — what a resolution above your plan allowance actually costs — is not stated on Zendesk's own pricing page; the ~$1.20-$1.50 figure on that listing is a third-party estimate, not a published one.

The free-tier split tracks this same divide closely: [Clay](/clay), [Manus](/manus), [ElevenLabs Conversational AI](/elevenlabs-conversational-ai) and [Genspark](/genspark) — 4 of these 22 — let you run real usage at $0 before paying anything, and all four sit in the public-rate group. None of the sales-led, custom-quoted tools offers any free tier or self-serve path at all.

By category, the filter bites hardest in [support agents](/category/support-agents): 8 of 9 bill by usage, the highest share of any category here, and support is where the outcome-based, sales-quoted pattern concentrates. [Voice agents](/category/voice-agents) contribute the same number of entries from a larger pool, 8 of 12 — per-minute billing is close to the category default there. [Research & data agents](/category/research-agents) add 4 of 11, and [coding agents](/category/coding-agents) 2 of 25.

| Name | What it is | Why it's here | Pricing | API |
| --- | --- | --- | --- | --- |
| [Vapi](https://theagentsindex.com/vapi.md) | The developer platform for voice AI agents, bring your own model, voice and telephony, and Vapi orchestrates the call. | $0.05/min pay-as-you-go orchestration fee, with the underlying model, voice and telephony billed through on top; about $10 in free trial credits, no permanent free tier. | — | Yes |
| [Retell AI](https://theagentsindex.com/retell-ai.md) | Build phone-call voice agents on Retell's turn-taking layer, plug in your own LLM, voice, and telephony, ship over voice, chat, or SMS. | $0.07-$0.31/min pay-as-you-go, with $10 in free credits and 20 free concurrent calls to start, but no permanent free tier. | — | Yes |
| [Bland AI](https://theagentsindex.com/bland-ai.md) | Enterprise voice AI that owns its own model stack, priced all-in per minute with no model pass-throughs. | $0.14/min with a $0 platform fee and no card required to start, but no permanent free tier once the starting credits run out. | — | Yes |
| [ElevenLabs Conversational AI](https://theagentsindex.com/elevenlabs-conversational-ai.md) | The voice-quality specialist’s own agent layer, 10,000+ voices and instant cloning, with a bring-your-own-LLM brain. | The one usage-based voice tool here with an ongoing free tier: 15 call minutes/month at $0, then LLM and telephony usage billed separately on top. | — | Yes |
| [Sierra](https://theagentsindex.com/sierra.md) | Bret Taylor’s enterprise agent platform, branded AI agents that resolve customer issues end to end, priced on outcomes. | Priced entirely on outcomes, a pre-negotiated rate per resolved issue, but that rate is never published; Sierra states pricing as 'Custom' with no self-serve option or free tier at all. | — | — |
| [Fin](https://theagentsindex.com/intercom-fin.md) | Intercom’s customer-service AI agent, resolves tickets omnichannel over any helpdesk, billed per resolution. | $0.99 per resolution with a 50-resolution monthly minimum; qualifications and handoffs are billed separately at $9.99 each. | — | Yes |
| [Decagon](https://theagentsindex.com/decagon.md) | Enterprise AI customer service agents for chat, voice and email. $4.5B valuation, custom contracts, no public price. | No published price list; a third-party cost analysis estimates roughly $50K/year platform fee plus $0.99/conversation and $0.50/resolution, not an official Decagon figure. | — | No |
| [Crescendo](https://theagentsindex.com/crescendo.md) | Fully-managed, AI-native CX platform that replaces your entire support stack with agents, QA and human staffing, billed per resolved outcome. | Priced per resolution from $2.99 with a 'Total Outcome Guarantee' (no charge for unresolved conversations), but no public self-serve tier; engagement starts with a paid pilot. | — | No |
| [Cresta](https://theagentsindex.com/cresta.md) | Unifies an autonomous AI Agent, real-time human Agent Assist and Conversation Intelligence on one contact-center platform, enterprise, custom-quoted, no self-serve tier. | An AWS Marketplace baseline of $150,000/year for one channel (chat or voice), then a published overage rate of $1.20/chat or $1.50/call beyond the contracted volume. | — | Yes |
| [Ada](https://theagentsindex.com/ada.md) | A Toronto-founded AI customer service agent platform from 2016. Sales-led, custom-quoted, with a Reasoning Engine, MCP server and a public V2 REST API that newer rivals in the category do not all publish. | No published price list; Vendr purchase-data analysis puts the median contract at $73,500/year, with independent estimates ranging roughly $30K to $300K+/year by volume. | — | Yes |
| [PolyAI](https://theagentsindex.com/polyai.md) | The pre-LLM-era enterprise voice AI incumbent ($750M-valued, 200+ enterprise customers): sales-led only again after a May 2026 self-serve tier went away. | Per-minute pricing bundling tuning, maintenance and support, per PolyAI's own pricing page, but the rate itself is never published; third-party estimates put typical contracts around $150,000/year and up. | — | Yes |
| [Synthflow AI](https://theagentsindex.com/synthflow-ai.md) | A no-code voice AI platform for enterprise phone agents — a visual Flow Designer and a structured Build-Evaluate-Launch-Learn deployment cycle, billed by the minute in separate components. | $0.11-$0.16/min typical, billed as three separate components (voice engine, LLM, telephony); no permanent free tier. | — | Yes |
| [Genspark](https://theagentsindex.com/genspark.md) | One web workspace for research, files, slides, code, media and agent-run tasks, backed by a persistent context layer. | Free tier included: about 100 credits/day refreshed daily with no card required, before its Super Agent workspace bills further usage. | — | Yes |
| [Microsoft Fabric Data Agent](https://theagentsindex.com/microsoft-fabric-data-agent.md) | A generally available, governed natural-language Q&A agent built into Microsoft Fabric. It turns plain-English questions into SQL, DAX, KQL or GQL over your own lakehouse, warehouse, Power BI, mirrored and graph data, run as the requesting user, permission-scoped and read-only. | Billed through Fabric's shared compute capacity rather than a per-query fee — the smallest capacity a data agent can run on costs roughly $0.36/hr (~$263/mo always-on), shared with every other Fabric workload on that capacity. | — | Yes |
| [Webhound](https://theagentsindex.com/webhound.md) | An AI research agent priced by the dollar, not a subscription. Set a budget, get a cited Report or a sourced Dataset, and trigger it from inside Claude, Codex or Cursor via a native MCP server. | Priced per dollar of compute with no subscription: roughly $1 for a narrow check, $5 for a standard cited report, scaling up from there; a one-time $5 free credit for new accounts. | — | Yes |
| [Zendesk AI Agents](https://theagentsindex.com/zendesk-ai-agents.md) | The customer-facing AI agent built into Zendesk's Suite helpdesk, now with Forethought's autonomous engine bundled in, billed per Verified Resolution on top of a per-seat plan. | Usage-metered on top of a per-seat plan starting at $55/agent/month, but the per-Verified-Resolution overage rate itself — roughly $1.20-$1.50 — is a third-party estimate; Zendesk does not publish the exact figure. | — | Yes |
| [OpenAI Realtime API](https://theagentsindex.com/openai-realtime-api.md) | OpenAI's own speech-to-speech model API for voice agents: WebRTC, WebSocket or native SIP telephony, billed purely on audio tokens with no platform fee. | Billed purely on audio tokens with no platform fee: $32/$64 per 1M tokens for the flagship model, working out to an independently estimated $0.06-$0.11/min in production once caching is in effect. | — | Yes |
| [Deepgram Voice Agent API](https://theagentsindex.com/deepgram-voice-agent-api.md) | The speech-infrastructure vendor's own voice-agent API. Managed STT, LLM and TTS over a single WebSocket, $4.50/hr flat, with 6 LLM providers and self-hosted or VPC deployment on the same product. | $0.075/min down to $0.050/min depending on which components you bring your own key for; a one-time $200 credit for new accounts, no permanent free tier. | — | Yes |
| [Auto-Respond](https://theagentsindex.com/auto-respond.md) | AI answering service that picks up calls and replies on Yelp, Thumbtack, Facebook, Instagram, and Google Local Services Ads, then books the job | A $10/month voice number plus $1/minute of talk time, alongside separate flat monthly fees per channel (Facebook, Thumbtack, Yelp) — a mixed usage-and-subscription structure, not a pure per-minute rate. | — | Yes |
| [Muse Code](https://theagentsindex.com/muse-code.md) | Meta's terminal coding agent — pay standard rate and it never trains on your code, or opt into Contributor pricing for a steep discount if you let it. | $1.25/M input and $4.25/M output tokens on the Standard tier, or a steep discount on the Contributor tier in exchange for letting it train on your code. | — | Yes |
| [Devin](https://theagentsindex.com/devin.md) | Devin is Cognition's autonomous coding agent: assign a task through Slack, CLI, or the web app, and it plans, edits, tests, and opens a pull request inside its own sandbox. | $500/month for 250 Agent Compute Units on the Team plan, or pay-as-you-go from $20/month on Core, billed by compute consumed rather than per seat; unlimited seats on every plan, and no confirmed ongoing free tier. | — | Yes |
| [Manus](https://theagentsindex.com/manus.md) | An autonomous general AI agent that plans and executes multi-step tasks end-to-end in its own cloud computer. | Free tier included: about 300 daily refresh credits, roughly one task a day, before paid usage kicks in. | — | Yes |

## How this list is built

A tool earns a place here if its own Published, sourced listing records `attributes.pricingModel` as `usage-based` — billed by a metered unit (minutes, resolutions, tokens, dollars of compute) rather than a flat per-seat subscription or one-time fee. Within that set, each entry's own `pricingTiers` data was read directly for whether the vendor publishes an actual per-unit number a buyer could use to estimate cost before signing up, or states the pricing model without the rate (`priceDisplay: 'Custom'`, sales-led only, no self-serve signup) — and, separately, whether an ongoing $0 tier exists (a limited free trial or one-time signup credit does not count as a free tier). Every number cited is the listing's own already-researched, sourced figure; where a figure is a third-party estimate rather than the vendor's own published number, the entry says so explicitly, exactly as the source listing does.