AI Agent Platforms Pricing Comparison: 15 Platforms Analyzed for 2026

AI agent platform pricing looks chaotic because vendors bill for four different things: tokens, messages, conversations, and outcomes. Once you sort the 15 platforms below by what the billable unit actually is, the comparison gets straightforward. This guide gives you the current numbers with sources, a worked cost example, and a checklist for choosing.

The four pricing models, not three tiers

Most comparisons group platforms by price. It is more useful to group them by billing unit, because the unit determines how your bill scales:

• Per token: you pay for the model compute your agent consumes. OpenAI and Anthropic APIs work this way. Cheapest per unit of work, but you build and maintain everything yourself.
• Per message or credit: you pay for each interaction the agent handles, bundled into packs. Microsoft Copilot Studio and Salesforce's Flex Credits work this way.
• Per conversation or resolution: you pay when the agent completes a unit of work. Salesforce Agentforce charges per conversation; Intercom Fin charges only when it resolves the issue.
• Per seat, per robot, or flat enterprise license: classic SaaS pricing applied to agents. Predictable, but you pay the same whether the agent works hard or idles.

Token pricing is the raw ingredient cost. Everything else is a finished product with a margin on top, and the sections below quantify that margin.

Developer APIs: the raw ingredient cost

Anthropic Claude API

Anthropic publishes list prices per million tokens, with no platform fee. As of October 2026, per Anthropic's published list prices:

• Claude Haiku 5.5: $0.10 input / $0.50 output per million tokens. The volume tier for classification, routing, and extraction.
• Claude Sonnet 5.5: $2.00 / $10.00. The default for most business agents.
• Claude Opus 5.5: $4.00 / $20.00. For complex multi-step reasoning where a wrong answer costs more than the tokens.
• Cost reducers: cache hits bill at roughly 10% of input price, and the Batch API halves both input and output for asynchronous work.

OpenAI API

OpenAI's pricing page lists per-million-token rates across several model families:

• gpt-6-luna: $0.10 input / $0.50 output. The high-volume tier.
• gpt-6.1-sol: $2.00 / $10.00. The mainstream workhorse.
• gpt-6-astra: $10.00 / $50.00. The top of the range.
• Legacy GPT-5 family: gpt-5 at $1.25 / $10, gpt-5-mini at $0.25 / $2, gpt-5-nano at $0.05 / $0.40, still available and widely used.

Cached input is billed at a steep discount on both APIs, which matters for agents that re-read the same system prompt or knowledge base on every call. If you want a deeper look at how these token, subscription, and outcome structures compare conceptually, our guide to AI agent pricing models walks through each one.

Google Vertex AI

Google bills per token for Gemini models plus the usual GCP infrastructure costs (storage, networking, platform services). Token rates are competitive with the two above, but your real bill depends heavily on architecture, so treat any single number with suspicion and model your own workload in the GCP calculator.

The old version of this post quoted API prices per thousand tokens. The industry has standardized on per-million pricing, and the absolute numbers have fallen by roughly an order of magnitude since early 2025. If you are budgeting from an older comparison article, your numbers are wrong in your favour.

Per-message platforms: Microsoft Copilot Studio

Copilot Studio is the default choice for Microsoft shops. Per current coverage of Copilot Studio's message-based pricing, the base plan is $200 per month for 25,000 messages, roughly $0.008 per additional message.

Two things catch buyers out:

• A "message" is not a conversation. A single user request can consume several messages as the agent reasons and calls tools. Define the unit with Microsoft before signing.
• Premium connectors and Power Platform licensing sit on top of the message pack.

Per-conversation platforms: Salesforce Agentforce

Salesforce's Agentforce pricing page now lists four buying models:

• Conversations: $2 per conversation.
• Flex Credits: $500 per 100,000 credits, where agent actions consume credits.
• Agentforce User License: $5 per user per month for employee-facing agents.
• Flat Fee Access: $125 per user per month for unmetered use.

The $2 conversation price is only the starting point. Agentforce assumes you are already paying for Salesforce CRM, and Data Cloud requirements plus implementation services add meaningfully to the first-year total. It makes sense when your support or sales process already lives in Salesforce; it is an expensive way to buy a standalone agent.

Per-outcome platforms: Intercom Fin and the support agents

Intercom's Fin is the cleanest example of outcome pricing: $0.99 per resolution, where a resolution means the customer confirms the issue is solved, does not follow up after Fin's answer, or Fin completes a workflow. You are charged once per conversation even if Fin answers multiple questions. Seat costs sit underneath: Intercom plans run roughly $39 to $139 per seat per month.

The interesting number is the margin over raw tokens. The same analysis measured real support conversations and found the underlying token cost was $0.0023 to $0.0575 per conversation depending on the model. Fin at $0.99 is therefore a 17x to 430x markup over ingredient cost. That is not a criticism. You are buying the helpdesk integration, the retrieval layer, the escalation logic, and the analytics. It is a reminder of what outcome pricing pays for.

Other support-agent platforms (Ada, Decagon, Moveworks) do not publish prices. Expect enterprise contracts, typically annual commitments negotiated against your conversation volume. Anyone quoting you a specific figure for these without a source is guessing.

The rest of the field, honestly

Several platforms from earlier versions of this comparison either no longer publish prices or never did. Rather than repeat unverified numbers:

• UiPath prices around robots, AI units, and orchestrator licensing; the mix changes often enough that you need a current quote.
• ServiceNow agents are licensed per module on enterprise contracts; nothing useful is public.
• Relevance AI and Botpress publish starting tiers for small teams, with real production usage moving to custom plans.
• Zapier's agent features bill inside Zapier's task-based plans, which penalize chatty multi-step agents.

Treat "contact sales" as a data point in itself: platforms that hide prices are selling to procurement departments, not to you.

Worked example: 10,000 support conversations a month

Assume 10,000 customer conversations per month, averaging 1,000 input tokens and 500 output tokens each (10M input, 5M output tokens total), and that an AI agent fully resolves 60% of them.

• Build on the Anthropic API, Sonnet 5.5: 10M x $2 + 5M x $10 = $70 per month in tokens. On Haiku 5.5 it drops to about $3.50. Add your engineering time, hosting, and the helpdesk glue.
• Build on the OpenAI API, gpt-6.1-sol: the same arithmetic, $70 per month, since the mainstream tiers are price-matched.
• Copilot Studio: at roughly 8 messages per conversation you need 80,000 messages. The $200 pack covers 25,000; the remaining 55,000 at ~$0.008 costs $440. Total around $640 per month.
• Intercom Fin: 6,000 resolutions x $0.99 = $5,940, plus seat licenses for your human agents.
• Salesforce Agentforce: 10,000 conversations x $2 = $20,000, before Data Cloud and implementation.

The spread between $70 and $20,000 for the same workload is the whole story of this market. Raw APIs are one to two orders of magnitude cheaper per unit of work; packaged platforms charge for the product around the model. Neither is wrong. Paying $5,940 for Fin is rational if it deflects 6,000 tickets that would otherwise cost you far more in human time. Paying $20,000 for Agentforce is rational if it replaces work inside a Salesforce instance you already run.

What actually drives your bill

• Volume profile: token pricing wins at low and spiky volume; flat and per-seat pricing wins when utilization is high and steady.
• Task complexity: a multi-step agent chain can consume 10x the tokens of a simple Q&A reply. Model complexity, not just conversation count.
• Retry and failure rates: per-token and per-message models bill you for failed attempts. Per-outcome models (Fin) do not, which transfers failure risk to the vendor.
• The definition of the billable unit: "message", "conversation", "resolution", and "credit" mean different things at every vendor, and every definition is negotiable before you sign.
• Ecosystem lock-in: Agentforce needs Salesforce, Copilot Studio wants Power Platform. The platform fee you already pay is part of the agent's real cost.

How to choose: a short checklist

• Have engineers and an unusual workflow? Build on the Anthropic or OpenAI APIs. The token economics above are hard for any packaged product to beat.
• Live in Microsoft 365 and need internal agents fast? Copilot Studio. Nail down the message definition first.
• Live in Salesforce with a high ticket volume? Agentforce, negotiated against committed volume.
• Run a support team and want AI deflection this quarter, without a build project? Intercom Fin or a comparable outcome-priced support agent.
• Want the outcome without owning the build at all? A managed agent service, where someone else runs the agent on API economics and charges you a flat monthly fee. Our own AI Employees service is structured exactly this way: an always-on agent on a fixed monthly price, built and maintained for you.
• Need something none of these fit? That is the case for a custom build, which starts to beat packaged pricing somewhere in the high thousands of conversations per month.

For a platform-by-platform cost breakdown from earlier this year, see our AI agent platforms cost comparison, and for a broader view of the platform landscape our platform pricing guide covers selection criteria beyond price.

Frequently asked questions

What is the cheapest way to run an AI agent?

Directly on a model API. At October 2026 list prices, a typical support conversation costs between $0.0023 and $0.0575 in tokens depending on the model, so even 100,000 conversations a month is a four-figure token bill. The catch is that you pay for everything around the model: engineering, hosting, monitoring, and integrations.

Why is Salesforce Agentforce so much more expensive per conversation than the APIs?

Because the $2 buys the finished product: the CRM integration, the deployment, the guardrails, and the vendor's margin. The underlying model tokens for a conversation cost a few cents. Whether the premium is worth it depends on how much of the surrounding work you would otherwise build yourself.

Is per-resolution pricing like Intercom Fin's a good deal?

It aligns cost with success, since you only pay when the AI actually resolves the issue, and failed attempts are free. At $0.99 per resolution it is expensive per unit compared to tokens, but cheap compared to a human-handled ticket. Watch the seat costs underneath and the vendor's definition of "resolved".

Should I wait for prices to fall before committing?

Token prices have fallen roughly an order of magnitude since early 2025, and the trend continues. But the model API is usually the smallest line in an agent budget; engineering and integration dominate. Waiting for cheaper tokens rarely changes the decision.

How do I compare vendors whose billing units all differ?

Convert everything to cost per resolved task in your own workload. Run a two-week pilot with each shortlisted platform on the same real conversations, count what each one billed, and divide by the number of tasks actually completed. That single number cuts through every pricing model.

Where to go from here

If the build-it-yourself economics appeal but the build itself does not, that is exactly the gap our AI Employees service fills: an always-on agent on API economics, run for you at a fixed monthly price. Talk to us about what your workload would actually cost.