.0097 to .33 across Agentforce, Intercom Fin, Copilot Studio and LLM APIs AI agent pricing comparison by Axis Intelligence Research: platform meters median AI Agent Cost 2026
By Axis Intelligence Research
Co-authors: Elena Rodriguez & Sarah Mitchell | Last updated: October 3, 2026 | License: CC BY 4.0
Prices verified on October 2, 2026. Every list price on this page was read directly from the vendor’s live pricing page or official documentation on that date. Where a vendor does not publish a price, we say so instead of estimating one.
Axis Intelligence Research finds that the same resolved customer conversation costs between $0.0097 and $3.33 at October 2026 list prices, a 344x spread, depending only on which billing meter an AI agent runs on. The median platform meter charges $0.99 per resolution; the median raw model API charges $0.09.
Quick Answer: How Much Does an AI Agent Cost in 2026?
An AI agent costs $0.99 per resolved conversation at the median packaged platform (Intercom Fin’s published outcome price, which is also the median of the five platform meters we priced) and about $0.09 per resolution in raw model tokens on a mid-tier model, according to the Axis Agent Resolution Cost Index (ARCI). The packaged-platform premium over raw tokens is roughly 10.8x. Seats, data-platform credits, and build costs sit on top.
Key Findings
- Axis Intelligence Research’s ARCI puts the median AI agent platform at $0.99 per resolved conversation and the median model API at $0.0917, as of October 2, 2026.
- Salesforce lists Agentforce at $0.10 per standard action on Flex Credits ($500 per 100,000 credits, 20 credits per action) and $2 per conversation on its Conversations meter, per its pricing page on October 2, 2026.
- Intercom prices Fin at $0.99 per outcome, and its pricing page counts a resolution when no further help is requested after Fin’s last answer.
- At the 2025 BLS median wage of $21.53 per hour, a $0.99 Fin outcome equals 2.76 minutes of a customer service representative’s pay; a $2 Agentforce conversation at a 60% resolution rate equals 9.29 minutes.
- Axis Intelligence Research estimates a 10,000-conversation-per-month support agent costs $71,280 a year on Fin’s meter versus $6,600 a year in Claude Sonnet 5.5 tokens, leaving $64,680 a year of headroom for anyone building in-house.
The Axis Agent Resolution Cost Index (ARCI): Why One Number Instead of Five Meters
Vendors bill AI agents in five incompatible units: per action, per conversation, per outcome, per credit, and per million tokens. None of their pricing pages lets a buyer compare across those units. So Axis Intelligence Research built one.
The Axis Agent Resolution Cost Index (ARCI) is the list-price cost of one resolved customer conversation, normalized across billing models. It answers the only question a support or operations leader actually budgets against: what does it cost when the agent closes a ticket without a human?
How ARCI is calculated. We define a reference conversation and price it on every meter at published list price:
- Platform meters: the reference conversation uses two knowledge answers and two system actions (look up an order, update a record). That is four Agentforce actions, or 14 Copilot Credits (2 generative answers at 2 credits each, 2 agent actions at 5 credits each, per Microsoft’s billing-rate table).
- Model APIs: the reference conversation consumes 20,000 input tokens and 1,500 output tokens. Agent loops re-send context on every tool call, so input dominates.
- Resolution adjustment: meters that bill every conversation are divided by an assumed 60% resolution rate, because the unresolved 40% still gets billed. Meters that bill only on outcome (Intercom Fin, Salesforce Help Agent Resolutions) are not divided.
Formula: ARCI = (list-price cost of one reference conversation) ÷ (resolution rate), with per-outcome meters taken at face value. Every input is a row in the downloadable dataset.
The 60% resolution rate and the token profile are Axis assumptions, not vendor data. We hold them constant so the comparison isolates price structure. Change them and the ranking holds for most pairs; the dataset lets you re-run it at any rate.
AI Agent Pricing Table: Every Published Price, Verified October 2, 2026
Packaged AI Agent Platforms
| Vendor | Product / tier | List price | Billing unit | Verified | Source |
|---|---|---|---|---|---|
| Salesforce | Agentforce Flex Credits | $500 per 100,000 credits ($0.10 per standard action, $0.15 per voice action) | Per action | 2026-10-02 | Salesforce |
| Salesforce | Agentforce Conversations | $2.00 | Per conversation | 2026-10-02 | Salesforce |
| Salesforce | Help Agent Resolutions | $2.00 | Per resolution | 2026-10-02 | Salesforce |
| Salesforce | Agentforce add-ons (Sales, Service, Field Service) | $125 | Per user per month | 2026-10-02 | Salesforce |
| Salesforce | Agentforce Max Editions | From $550 (includes 2.75M Flex Credits per org per year) | Per user per month | 2026-10-02 | Salesforce |
| Salesforce | Agentforce User License | $5 (requires Flex Credits) | Per user per month | 2026-10-02 | Salesforce |
| Intercom | Fin AI Agent | $0.99 ($9.99 per sales qualification) | Per outcome | 2026-10-02 | Intercom |
| Intercom | Fin on another helpdesk | $0.99, 50-outcome monthly minimum | Per outcome | 2026-10-02 | Intercom |
| Intercom | Helpdesk seat with Fin | From $29 | Per seat per month | 2026-10-02 | Intercom |
| Microsoft | Copilot Studio capacity pack | $200 per 25,000 Copilot Credits ($0.008 per credit) | Per month | 2026-10-02 | Microsoft |
| Microsoft | Microsoft 365 Copilot | $30, paid yearly | Per user per month | 2026-10-02 | Microsoft |
| Zendesk | AI agents, automated resolutions | Not published | Per automated resolution | 2026-10-02 | Zendesk (no public rate card) |
| Intercom | Fin Voice | Custom pricing | Quote-based | 2026-10-02 | Intercom |
Foundation Model APIs (the “build your own agent” layer)
| Vendor | Model | Input (per 1M tokens) | Output (per 1M tokens) | Verified | Source |
|---|---|---|---|---|---|
| OpenAI | GPT-5.6 Luna | $0.20 | $1.20 | 2026-10-02 | OpenAI |
| Gemini 3.8 Flash | $0.75 through Dec 31, 2026; $1.50 from Jan 1, 2027 | $3.75; $7.50 from Jan 1, 2027 | 2026-10-02 | ||
| Anthropic | Claude Haiku 4.5 | $1 | $5 | 2026-10-02 | Anthropic |
| Anthropic | Claude Sonnet 5.5 | $2 | $10 | 2026-10-02 | Anthropic |
| OpenAI | GPT-5.6 Terra | $2 | $12 | 2026-10-02 | OpenAI |
| Gemini 3.1 Pro Preview (prompts ≤200k) | $2 | $12 | 2026-10-02 | ||
| Anthropic | Claude Opus 5.5 | $4 | $20 | 2026-10-02 | Anthropic |
| OpenAI | GPT-5.6 Sol | $5 | $30 | 2026-10-02 | OpenAI |
| Anthropic | Claude Fable 5.1 | $10 | $50 | 2026-10-02 | Anthropic |
Zendesk’s per-resolution rate is absent from its public pricing, and we found no primary document that states it. Rather than repeat third-party guesses, we leave the cell empty. Treat any single Zendesk number you see quoted as a negotiated figure, not a list price.
How Much Does an AI Agent Cost Per Resolution? The ARCI Ranking
| Rank (cheapest first) | Meter | Cost per reference conversation | ARCI: cost per resolved conversation | Wage-minute equivalent* |
|---|---|---|---|---|
| 1 | GPT-5.6 Luna (tokens) | $0.0058 | $0.0097 | 0.03 min |
| 2 | Gemini 3.8 Flash, 2026 promo rate (tokens) | $0.0206 | $0.0344 | 0.10 min |
| 3 | Claude Haiku 4.5 (tokens) | $0.0275 | $0.0458 | 0.13 min |
| 4 | Gemini 3.8 Flash, 2027 rate (tokens) | $0.04125 | $0.0687 | 0.19 min |
| 5 | Claude Sonnet 5.5 (tokens) | $0.0550 | $0.0917 | 0.26 min |
| 6 | GPT-5.6 Terra / Gemini 3.1 Pro Preview (tokens) | $0.0580 | $0.0967 | 0.27 min |
| 7 | Claude Opus 5.5 (tokens) | $0.1100 | $0.1833 | 0.51 min |
| 8 | Microsoft Copilot Studio (14 credits at pack rate) | $0.112 | $0.1867 | 0.52 min |
| 9 | GPT-5.6 Sol (tokens) | $0.1450 | $0.2417 | 0.67 min |
| 10 | Salesforce Agentforce Flex Credits (4 actions) | $0.40 | $0.6667 | 1.86 min |
| 11 | Intercom Fin (per outcome) | n/a | $0.99 | 2.76 min |
| 12 | Salesforce Help Agent Resolutions | n/a | $2.00 | 5.57 min |
| 13 | Salesforce Agentforce Conversations | $2.00 | $3.33 | 9.29 min |
*Wage-minute equivalent = ARCI ÷ ($21.53 ÷ 60), using the BLS median hourly wage for customer service representatives for May 2025. Wage only, before benefits and overhead. Source: Axis Intelligence Research calculation from vendor list prices verified October 2, 2026.
Read the table as two markets wearing one label. The token rows price inference. The platform rows (Copilot Studio, Flex Credits, Fin, Help Agent, Conversations) price a product: the orchestration, the helpdesk integration, the testing harness, the analytics, and the vendor’s willingness to own the failure. A $0.99 outcome is not 100 times “more expensive” than a $0.0097 token bill. It is a different line on the P&L.
The useful comparison is inside each market. Among packaged platforms, the spread from Copilot Studio ($0.19) to Agentforce Conversations ($3.33) is 17.9x for the same reference conversation. That gap is the procurement opportunity, and it comes almost entirely from billing structure rather than capability.
AI Agent Pricing Models Compared: Per Action, Per Conversation, Per Outcome, Per Token
Per action (Salesforce Flex Credits, Microsoft Copilot Credits). Cost tracks how much work the agent does. A two-action order-status check is cheap; a twelve-step troubleshooting flow is not. Salesforce’s own worked example prices a three-action case-management task at $0.30 and a five-action field-service scheduling task at $0.50. The risk sits with the buyer: a badly designed agent that loops through extra tool calls bills for every loop.
Per conversation (Salesforce Conversations at $2). Flat and predictable per contact, but it charges the same for a conversation the agent abandons as for one it closes. At a 60% resolution rate, the effective price per resolution is $3.33. Salesforce also notes that Flex Credits and Conversations cannot run in the same org, so this is a one-way architectural choice per org, not a per-use-case toggle.
Per outcome (Intercom Fin at $0.99, Salesforce Help Agent at $2). The vendor carries the resolution risk, which is why outcome pricing carries the highest per-unit sticker. The fine print matters more here than anywhere: Intercom’s published definition counts a resolution when the customer asks for no further help after Fin’s last answer. A customer who gives up and leaves silently can therefore register as a billable outcome. Audit a sample of “resolved” conversations in month one.
Per token (OpenAI, Anthropic, Google APIs). The cheapest unit and the most work. You pay for inference only; retrieval, guardrails, evaluation, hosting, and on-call engineering are yours. Two token traps catch first-time builders. Anthropic’s pricing page notes that its newer tokenizer (Claude 4.7 and later) produces roughly 30% more tokens for the same text, so per-token price comparisons across generations understate real cost differences. And Google has published a scheduled doubling of Gemini 3.8 Flash list prices on January 1, 2027, which our ARCI ranking shows as two separate rows.
Fully-Loaded AI Agent TCO: 1, 3, and 5 Years at 10,000 Conversations a Month
To translate unit prices into budget lines, Axis Intelligence Research modeled a mid-market support operation: 10,000 customer conversations per month, 60% resolved by the agent (6,000 resolutions), list prices held flat except where the vendor has published a future change.
Formula: Annual meter cost = monthly billable units × list price × 12. Multi-year = annual × years.
| Meter | Billable units per month | Year 1 | 3 years | 5 years |
|---|---|---|---|---|
| Salesforce Agentforce Conversations | 10,000 conversations × $2 | $240,000 | $720,000 | $1,200,000 |
| Salesforce Help Agent Resolutions | 6,000 resolutions × $2 | $144,000 | $432,000 | $720,000 |
| Intercom Fin | 6,000 outcomes × $0.99 | $71,280 | $213,840 | $356,400 |
| Salesforce Agentforce Flex Credits | 40,000 actions × $0.10 | $48,000 | $144,000 | $240,000 |
| GPT-5.6 Sol API | 10,000 conversations × $0.145 | $17,400 | $52,200 | $87,000 |
| Microsoft Copilot Studio | 140,000 credits = 6 packs × $200 | $14,400 | $43,200 | $72,000 |
| Claude Sonnet 5.5 API | 10,000 conversations × $0.055 | $6,600 | $19,800 | $33,000 |
| Gemini 3.8 Flash API (2027 rate) | 10,000 conversations × $0.04125 | $4,950 | $14,850 | $24,750 |
| Claude Haiku 4.5 API | 10,000 conversations × $0.0275 | $3,300 | $9,900 | $16,500 |
Source: Axis Intelligence Research estimates from vendor list prices verified October 2, 2026. Meter cost only.
The meter is not the TCO. Here is what the deck usually leaves out, by layer:
- Seats that ride alongside the meter. Intercom helpdesk seats start at $29 per seat per month when you run Fin inside Intercom; Fin on another helpdesk has no seat charge but carries a 50-outcome monthly minimum. Salesforce add-ons run $125 per user per month for unmetered employee use.
- Data platform consumption. Salesforce’s worked examples explicitly exclude Data 360 credits and other consumption services. If your agent grounds on unified customer data, budget that meter separately.
- Grounding and search fees. Microsoft bills tenant graph grounding at 10 Copilot Credits per use, five times a generative answer, which would take our 14-credit reference conversation to 24 credits. Anthropic and OpenAI each charge $10 per 1,000 web searches; Google charges $14 per 1,000 grounded searches after 5,000 free per month.
- Build and run for the DIY path. The token-only rows exclude engineering, hosting, evaluation, and monitoring. That is the real comparison: at 10,000 conversations a month, Fin’s meter costs $64,680 a year more than Claude Sonnet 5.5 tokens. That $64,680 is the annual budget an in-house build must stay under, for everything except tokens, to win on cost.
Elena Rodriguez’s read: the ROI deck almost always runs the vendor’s resolution rate, not yours. Run the table above at 40% and the per-conversation meters get worse fast. Salesforce Conversations rises to $5.00 per resolution, while Fin’s outcome price stays at $0.99 because unresolved conversations aren’t billed. The meter you pick is a bet on your own containment rate, and most teams have never measured theirs.
Is an AI Agent Cheaper Than a Human Agent?
Per resolved contact, at list price, usually yes, and the wage-minute equivalent shows by how much. The BLS puts the May 2025 median wage for customer service representatives at $21.53 an hour, or about 35.9 cents per minute. A $0.99 Fin outcome buys 2.76 minutes of that median wage. A Copilot Studio resolution at pack rates buys about half a minute.
So the break-even question becomes concrete: if your human agents resolve the same contact type in less time than the wage-minute equivalent, the AI meter costs more than the labor it replaces. For Salesforce Conversations at a 60% resolution rate, that threshold is 9.29 minutes of median wage. For simple password resets and order-status checks that humans already clear quickly, a per-conversation meter can lose that comparison. For multi-step returns and billing disputes, it wins easily.
Two caveats keep this honest. The BLS figure is wage only; benefits, management, and facilities push the true human cost per minute higher, which favors the AI side. And the AI meter covers only the resolved share; the 40% that escalates still lands on a human.
What Moves the Price of an AI Agent?
Resolution rate. The single largest lever for per-conversation and per-action meters. At 60% resolution, Agentforce Conversations costs $3.33 per resolution; at 80%, $2.50; at 40%, $5.00. Outcome meters are immune by design.
Actions per conversation. On Flex Credits, every action is $0.10. Salesforce’s per-action meter beats its own $2 Conversations meter until an agent averages 20 actions per conversation, and beats the $2 Help Agent resolution price until it averages 12 actions per resolved conversation at a 60% resolution rate. Against Fin’s $0.99, Flex Credits stay cheaper only up to about 6 actions per conversation at the same 60% resolution rate.
Voice versus text. Salesforce bills Agentforce Voice actions at 30 credits ($0.15) versus 20 for text. Intercom prices Fin Voice on custom quotes. OpenAI’s realtime voice model lists audio input at $32 and audio output at $64 per million tokens, against $5 and $30 for GPT-5.6 Sol text.
Model tier. Inside one vendor, the jump from the small model to the flagship moves ARCI by 25x at OpenAI (Luna to Sol) and 4x at Anthropic (Haiku 4.5 to Opus 5.5). Routing easy turns to a small model is the cheapest optimization available.
Caching and batch. Anthropic bills cache hits at 10% of base input price on most models, and both Anthropic and OpenAI discount batch processing by 50%. Because agent loops re-send the same system prompt and tool definitions on every call, caching cuts the input-heavy reference conversation substantially.
Commitment structure. Microsoft’s capacity pack prices credits at $0.008 versus pay-as-you-go. Salesforce states unused Flex Credits do not roll over into the next subscription term. Prepaying saves money only if utilization stays high.
Who uses the agent. Microsoft’s billing documentation states that interactive use by authenticated Microsoft 365 Copilot users is included at no extra credit cost. An internal IT-helpdesk agent for licensed staff can have a near-zero marginal meter; the same agent on your public website bills every response.
How to Buy an AI Agent in 2026: Pricing Guidance by Scenario
You run Salesforce Service Cloud and your conversations are short. Model Flex Credits first. At fewer than 12 actions per resolved conversation, Flex undercuts both $2 meters at a 60% resolution rate. Ask your account executive for a Pre-Commit structure only after three months of Digital Wallet usage data.
You need predictable spend and can’t measure containment yet. An outcome meter (Fin at $0.99) shifts resolution risk to the vendor. Negotiate the resolution definition in writing, and require a sampling audit of silent-exit “resolutions.”
You live in Microsoft 365 and the agent is employee-facing. Check whether your users already hold Microsoft 365 Copilot licenses. If they do, interactive internal usage may not draw credits at all, and Copilot Studio becomes the lowest-ARCI packaged option in this study.
You have an engineering team and more than a few thousand conversations a month. Price the build against the headroom. At 10,000 conversations a month, the gap between Fin’s meter and Claude Sonnet 5.5 tokens is $64,680 a year; at 50,000 conversations, it scales linearly to $323,400. If one engineer plus hosting and evaluation tooling costs less than the headroom, building wins on cost.
You’re piloting voice. Get written quotes. Two of the three platform vendors in this study either price voice above text or price it only on request.
Disclosure: Some links on axis-intelligence.com may be affiliate or partner links. Rankings in this article are ordered strictly by computed ARCI value; no vendor paid for inclusion or placement, and commercial relationships carry zero weight.
AI Agent Cost: Questions Buyers Ask Before Signing
What is the cheapest way to run an AI customer service agent in 2026?
At list price, the lowest cost per resolved conversation in this study is GPT-5.6 Luna tokens at about $0.0097, per Axis Intelligence Research’s ARCI. Among packaged platforms, Microsoft Copilot Studio is lowest at about $0.19 per resolution at capacity-pack rates. Packaged platforms include orchestration and tooling that raw tokens do not.
Does Intercom Fin charge when the customer just leaves the chat?
It can. Intercom’s pricing page defines a billable resolution as one where no further help is requested after Fin’s last answer, at $0.99 per outcome. A customer who abandons the conversation without asking for more help can fall under that definition. Audit a sample of resolved conversations early.
Should I choose Agentforce Flex Credits or Conversations pricing?
Choose Flex Credits if your agent averages fewer than 20 actions per conversation. Flex bills $0.10 per standard action; Conversations bills a flat $2. Salesforce does not allow both models in the same org, and unused Flex Credits do not roll over between subscription terms.
Why does my AI agent cost more per resolution than the vendor’s price?
Because per-action and per-conversation meters bill unresolved conversations too. Divide your monthly meter cost by resolved conversations, not total conversations. At a 60% resolution rate, a $2 per-conversation price becomes $3.33 per resolution, according to Axis Intelligence Research.
How many minutes of human labor does one AI resolution cost?
Axis Intelligence Research’s wage-minute equivalent converts AI prices into BLS median wage time. A $0.99 Fin outcome equals 2.76 minutes of a customer service representative’s median $21.53 hourly wage; a Copilot Studio resolution equals about half a minute.
Is Copilot Studio free with Microsoft 365 Copilot?
For internal use, often yes. Microsoft’s billing documentation says interactive use of agents by authenticated Microsoft 365 Copilot users is included at no extra credit cost. Customer-facing agents and agents used by unlicensed people consume Copilot Credits at $200 per 25,000-credit pack.
Are Gemini API prices going up in 2027?
Yes, for Gemini 3.8 Flash. Google’s pricing page lists $0.75 input and $3.75 output per million tokens through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027. Budget multi-year agent projects at the 2027 rate.
What does it cost to build an AI agent instead of buying one?
The token bill is small; the build is not. At 10,000 conversations a month, Claude Sonnet 5.5 tokens cost about $6,600 a year versus $71,280 on Intercom Fin’s meter, per Axis Intelligence Research. The $64,680 difference is the yearly ceiling for engineering, hosting, and evaluation if building is to win on cost.
Why don’t Zendesk AI agent prices appear in the table?
Zendesk bills AI agents per automated resolution but does not publish the per-resolution rate on its public pricing. Axis Intelligence Research found no primary document stating it, so the rate is listed as not published rather than estimated from secondhand reports.
Methodology
Data collection. Axis Intelligence Research retrieved every list price on October 2, 2026, from the vendor’s own pricing page or official documentation: Salesforce’s Agentforce pricing page and Help article 004811240; Intercom’s Fin pricing page; Microsoft’s Copilot Studio plans page and Microsoft Learn billing-rate documentation; Anthropic’s Claude Platform pricing documentation; OpenAI’s API pricing page; and Google’s Gemini Developer API pricing page. The labor benchmark is the BLS Occupational Outlook Handbook median hourly wage for customer service representatives (May 2025 OEWS data). No secondary aggregator figures are used.
The ARCI formula. ARCI = list-price cost of one reference conversation ÷ resolution rate. Reference conversation: two knowledge answers plus two system actions on platform meters (4 Agentforce actions; 14 Copilot Credits); 20,000 input and 1,500 output tokens on model APIs. Resolution rate: 60%. Per-outcome meters are taken at face value. Copilot Credits are priced at the capacity-pack rate of $0.008 per credit ($200 ÷ 25,000).
Wage-minute equivalent (WME). WME = ARCI ÷ ($21.53 ÷ 60).
TCO model. 10,000 conversations per month; 6,000 resolutions; list prices held flat; Gemini 3.8 Flash at its published 2027 rate; Copilot Studio rounded up to whole packs (6 packs for 140,000 credits).
What ARCI does not price. Negotiated discounts, seats, data-platform credits, build and hosting costs, and quality differences between agents. ARCI compares price structures, not answer quality.
Download the dataset. ai-agent-cost-2026.csv (126 rows, CC BY 4.0). Every list price, every ARCI calculation, and every TCO line includes source organization, URL, retrieval date, and a method note.
Cite This Page
APA: Axis Intelligence Research, & Rodriguez, E. (2026, October 3). AI agent cost 2026: What one resolved conversation really costs. Axis Intelligence. https://axis-intelligence.com/ai-agent-cost/
MLA: Axis Intelligence Research, and Elena Rodriguez. “AI Agent Cost 2026: What One Resolved Conversation Really Costs.” Axis Intelligence, 3 Oct. 2026, axis-intelligence.com/ai-agent-cost/.
Chicago: Axis Intelligence Research and Elena Rodriguez. “AI Agent Cost 2026: What One Resolved Conversation Really Costs.” Axis Intelligence, October 3, 2026. https://axis-intelligence.com/ai-agent-cost/.
Embeddable citation: According to Axis Intelligence Research’s Agent Resolution Cost Index (ARCI), the median AI agent platform charged $0.99 per resolved conversation in October 2026, about 10.8 times the median raw model API cost (Axis Intelligence Research, October 2026, CC BY 4.0).
Related Axis Intelligence Research
- AI agents statistics: market size, adoption, and the deployment gap
- Agentic AI statistics: market growth and enterprise impact data
- AI assistant statistics: adoption, usage, and the productivity gap
- AI statistics 2026: market size, adoption, and the value gap
- Agentic commerce statistics: AI traffic and the trust gap
- OpenAI API Pricing
- Claude API Pricing
- LLM API Pricing