How much does a custom AI agent cost an SMB? What drives the price
A custom AI agent's price depends on four things: the model, the number of connected tools, your CRM and human review. Here is what each one changes on your bill.
4 min read
The SynAI team
Key points
- No vendor's official page publishes the cost of building a custom AI agent; OpenAI and Anthropic publish only the usage price of their models.
- The cost of an AI agent varies with the model chosen, the number of tools the agent operates and the API limits of the company's CRM subscription.
- In Quebec, Law 25 requires that a person subject to a fully automated decision be able to submit observations to an employee who can review that decision.
Is there a standard price for a custom AI agent?
No. No vendor's official page gives the cost of building a custom agent, so we don't quote a range. Vendors publish usage prices instead. A budget therefore has two parts: the build work, then monthly fees for the model, the orchestration platform and sometimes the CRM.
How much does an agent's AI model cost?
At OpenAI, the largest model costs 100 times more than the smallest. Pricing is per million tokens sent to the model (input) and generated by it (output).
| Model | Input | Output |
|---|---|---|
| gpt-6-luna (OpenAI) | $0.10 | $0.50 |
| gpt-6-astra (OpenAI) | $10.00 | $50.00 |
| Claude Haiku 5.5 (Anthropic) | $0.10 | $0.50 |
| Claude Sonnet 5 (Anthropic) | $2.00 | $10.00 |
| Claude Opus 5.5 (Anthropic) | $4.00 | $20.00 |
Model
gpt-6-luna (OpenAI)
- Input
- $0.10
- Output
- $0.50
gpt-6-astra (OpenAI)
- Input
- $10.00
- Output
- $50.00
Claude Haiku 5.5 (Anthropic)
- Input
- $0.10
- Output
- $0.50
Claude Sonnet 5 (Anthropic)
- Input
- $2.00
- Output
- $10.00
Claude Opus 5.5 (Anthropic)
- Input
- $4.00
- Output
- $20.00
USD per million tokens, standard rate, short context, as of October 13, 2026. Haiku 5.5: prompts up to 100,000 tokens.
Batch processing, for tasks that can wait, costs half the standard rate at Anthropic, and for gpt-6-astra and gpt-6.1-sol at OpenAI. At Anthropic, a cache hit costs 10% of the standard input price.
Anthropic recommends the simplest possible solution, which may mean not building an agent at all. OpenAI advises prototyping with the most capable model, then trying smaller ones. Our article on Claude Opus 5.5 compares its prices with Opus 5.
Why does the number of connected tools change the price?
Every tool the agent operates has to be documented, tested and controlled. According to OpenAI, a tool either reads data, such as a CRM or a PDF, or takes action: sending an email, updating a customer record, handing a ticket to a person.
- Overlap. OpenAI says the issue isn't only how many tools there are but how similar they are: some agents handle more than 15 distinct tools, while others struggle with fewer than 10 overlapping ones.
- Risk. OpenAI recommends rating each tool low, medium or high based on read-only versus write access, reversibility, permissions and financial impact.
- Testing. Anthropic recommends building these interfaces carefully, with thorough documentation for each tool and testing.
How does CRM integration affect the budget?
Your CRM's API limits depend on your subscription, and the agent uses them up. At HubSpot, an app that exceeds its limit gets a 429 error on every subsequent call.
| CRM | Edition | Usage cap |
|---|---|---|
| HubSpot | Free and Starter | 250,000 calls |
| HubSpot | Professional | 625,000 calls |
| HubSpot | Enterprise | 1,000,000 calls |
| Zoho CRM | Standard | 50,000 credits + 250 per licence (max. 100,000) |
| Zoho CRM | Professional | 500 credits per licence (max. 3,000,000) |
CRM
HubSpot
- Edition
- Free and Starter
- Usage cap
- 250,000 calls
HubSpot
- Edition
- Professional
- Usage cap
- 625,000 calls
HubSpot
- Edition
- Enterprise
- Usage cap
- 1,000,000 calls
Zoho CRM
- Edition
- Standard
- Usage cap
- 50,000 credits + 250 per licence (max. 100,000)
Zoho CRM
- Edition
- Professional
- Usage cap
- 500 credits per licence (max. 3,000,000)
HubSpot: private apps, per account per day. Zoho: rolling 24-hour window. As of October 13, 2026.
The platform that connects the tools has a price too. n8n bills by monthly workflow executions: €20 a month for 2,500 executions (Starter) and €50 for 10,000 (Pro), billed annually.
Why does human review weigh on the budget?
You have to plan when the agent stops and who approves. OpenAI recommends human oversight for sensitive or irreversible actions: cancelling an order, authorizing a large refund, making a payment. Anthropic describes checkpoints and stopping conditions, such as a maximum number of iterations. Our Hermes Agent vs OpenClaw comparison covers the guardrails to set up.
Frequently asked questions
- Are model prices in Canadian dollars?
- No. OpenAI and Anthropic list their prices in US dollars, and n8n in euros. The exchange rate therefore moves your bill.
- Where does the agent process data?
- At Anthropic, requiring US-only processing multiplies token prices by 1.1; that is US hosting, not Canadian. In Quebec, sending personal information outside the province first requires a privacy impact assessment (Law 25, s. 17). Our Claude on a VPS article explains how to host an agent on a server.
- Can the project be subsidized?
- Our article on AI grants for Quebec SMBs lists the programs worth checking.
Sources
- 1OpenAI · API Pricing
- 2Anthropic · Pricing
- 3Anthropic · Building effective agents
- 4OpenAI · A practical guide to building AI agents
- 5HubSpot · Usage guidelines
- 6Zoho · CRM API Limits
- 7n8n · Pricing
- 8LégisQuébec · Act respecting the protection of personal information in the private sector, P-39.1 (in French)
- 9CNIL · GDPR, Chapter III: rights of the data subject (in French)
