This article is part of our AI Agent Guides. See how models compare this week on the AI Agent Leaderboard, a ranking of rankings refreshed every week.
Pick per-seat pricing when people work alongside the agent, and usage pricing when you are building it yourself. Pick per-resolution or per-outcome pricing when the agent finishes customer work alone. Most buyers end up with a mix, so model your volume before you sign.
Prices here come from vendor pages checked on 2 November 2026. They change, so confirm before you commit. Pricing tells you what you pay, not how well an agent works, so check model quality separately on the current leaderboard.
Disclosure: eSEOspace is a member of the Claude Partner Network.
At a glance
| Model | You pay for | Example | Biggest risk |
|---|---|---|---|
| Per seat | Each person | Claude Team, ChatGPT Business | Hidden usage limits, idle seats |
| Usage or tokens | Tokens consumed | Claude, OpenAI, Gemini APIs | Bill grows with loops and retries |
| Credits | Metered actions | Agentforce Flex Credits, Copilot Studio | Hard to translate into tasks |
| Per conversation | Each conversation | Agentforce Conversations | Pay even when the agent fails |
| Per resolution | Each resolved request | Zendesk automated resolutions | Definition of resolved |
| Per outcome | Defined results | Fin, Sierra | Definition, and paying for weak outcomes |
Per seat
You pay a fixed price for each person. It suits tools that staff use directly, such as coding agents or research assistants.
Anthropic lists Team Standard at $20 per seat on annual billing or $25 monthly, and Premium at $100 or $125. OpenAI's Codex page lists Business at $20 or $25. Google lists Gemini Enterprise from $21 or $30 per seat, depending on edition.
The trap is that seats are not unlimited. Anthropic's Team allowance resets on five-hour and weekly windows, and OpenAI says its message figures are estimates, not fixed limits. Ask what happens when a heavy user runs out.
Some developer tools blend the two models.
GitHub Copilot's plans pair monthly prices with AI credits consumed by agent mode and code review. Cursor and Devin sell individual tiers up to $200 a month. Devin's page says Pro users can buy extra usage at API pricing.
Usage or tokens
You pay for tokens, the small chunks of text a model reads and writes. This is the normal model when you build your own agent.
It is fair and transparent, since you pay only for what runs. As of 2 November 2026, Anthropic, OpenAI and Google all publish token rates and batch discounts.
The risk is variance. Agents loop, retry and carry long context, so spend can jump with no change in your customer volume.
Set workspace spend limits and alerts on day one. Our cost-per-task guide shows how to forecast this.
Credits
Credits are a usage model with an extra layer of translation. You buy a pool, and each action draws it down.
Salesforce's Agentforce pricing lists Flex Credits at $500 per 100,000. Its worked example charges 40 credits, or $0.20, for a two-action request. At 20 requests a day for 30 days, that comes to $120 a month.
Microsoft Copilot Studio also meters in credits. It lists 1 credit for a classic answer, 5 for an agent action and 10 for tenant graph grounding. An interaction can combine several, so one customer question can use many credits.
The upside is flexibility across use cases. The downside is that credits are hard to compare between vendors. Always convert them to dollars per finished task.
Per conversation
You pay a flat fee for each conversation. Agentforce lists this option at $2 per conversation, in addition to its credit option.
This is easy to budget if you know your volume. However, you may pay whether or not the customer's problem was solved. Compare it against your current contact cost, and against outcome pricing on the same volume. Use a real month of tickets, including your busiest week.
Per resolution
You pay only when the agent resolves the request. Zendesk bills through automated resolutions, defined as requests successfully resolved without escalation to a human agent. Each plan includes an allowance, and you can commit to more or pay as you go.
This shifts some risk to the vendor, which is appealing. But notice how much depends on one word. Ask how the vendor decides that a request was resolved.
Per outcome
Outcome pricing broadens the idea to other results, such as a saved cancellation or a qualified lead. Sierra's post on outcome-based pricing, from December 2024, says that in most cases an unresolved conversation carries no charge. Sierra does not publish a rate card there, so expect to talk to sales.
Fin's pricing page gives clear numbers. It lists $0.99 per outcome for resolutions, procedure handoffs, disqualifications and self-serve routings, and $9.99 for qualifications. You are not charged when a conversation simply passes to your team without an outcome.
Read Fin's definition closely: a resolution means no further help is requested after Fin's last answer.
Our take: that is clear but generous to the seller. A customer who gives up and leaves may look the same as one who was helped. This is a general issue with outcome pricing, not just Fin.
Contract traps by model
Whichever you choose, these questions will save you money.
| Question | Why it matters |
|---|---|
| What counts as a billable event? | One vague word can change the bill by a lot |
| How is customer silence treated? | Silence can look like success |
| Can a customer's reopened request be refunded? | Otherwise you pay twice for one problem |
| What happens at a limit? | The agent may stop, slow down or bill overage |
| Are there spend caps and alerts? | Runaway loops need a hard stop |
| Is there a platform fee on top? | Per-unit prices may sit on a base plan |
| Can you audit billed events? | You should be able to sample and dispute them |
| When can rates change? | Google already lists increases for some models |
How to compare two quotes
Convert every quote to one number. Use your last 90 days of real volume.
- Estimate how many tasks the agent would attempt each month.
- Estimate how many it would finish, using a trial rather than a sales claim.
- Apply each vendor's price, including base fees and overages.
- Add staff time for review and escalation.
- Divide by tasks finished.
Here is a made-up example. Assume 1,000 support requests a month and a 50% resolution rate. At $0.99 per outcome, that is about $495, while a $2 per conversation plan would be $2,000.
Those numbers do not rank the vendors. They show how sensitive the result is to your resolution rate and to what each vendor counts. So run the math on your own volume, and test quality with our agent scorecard.
Which model suits which job
This is our editorial view, based on how each pricing model works rather than on any study. Use it as a first guess, then check it against your own volume and your own trial results.
| Job | Usual fit | What to watch |
|---|---|---|
| Developer coding help | Per seat | Usage ceilings for heavy users |
| Custom internal agent | Usage or tokens | Spend caps, loops |
| Customer support | Per resolution or outcome | What counts as resolved |
| Lead qualification | Per outcome | Lead quality, not just count |
| Work inside a CRM | Credits or per user | Converting credits into dollars per task |
Benchmarks do not price products
Leaderboards measure models, while vendors price products. A product may use a mix of models or change them without notice. So use the leaderboard methodology to understand what benchmark numbers mean.
Then ask each vendor which models power the product. Ask also whether a model change would alter your pricing or results.
Which should you choose
Choose per seat if people drive the agent and your use is steady. Choose usage if you are building your own agent and can watch spend closely. Our build vs buy guide helps with that decision.
Choose per resolution or outcome if the agent replaces a clear, countable task like support tickets. Insist on tight definitions and audit rights.
Choose credits or per conversation if you are already inside that vendor's platform, and convert them to cost per task first. If you need a second opinion on a quote, you can contact us.
Frequently asked questions
Which AI agent pricing model is cheapest?
None is cheapest in general. Low volume often favors usage or outcome pricing, while heavy daily use can favor seats. Run your own volume through each model before you decide.
Is outcome-based pricing risk free?
No. It moves some risk to the vendor, but you still carry risk from weak definitions and bad customer experiences. Sample billed outcomes every month and check them yourself.
Why do vendors use credits?
Credits let one pool cover many actions and products. They also make direct price comparison harder for you. Ask the vendor for a worked example that matches your real task.
Should I sign an annual contract?
Be careful. Prices, models and products are changing quickly, and annual discounts on seats are modest in the examples above. Start monthly or with a short pilot if you can, then commit once you have data.
