What an AI Agent Costs to Run: Tokens, Keys and a Budget per Agent for Odoo and MYOB Acumatica

Updated 5 October 2026: Odoo licence figures now use the Australian pricelist that took effect on 24 September 2026 (Custom A$64.00 for the first year, then A$80.00 at list; Standard unchanged).

As at 15 September 2026, published hosted-model prices run from about US$0.20 to US$10 per million input tokens and US$1.20 to US$50 per million output tokens, so one drafts-first agent handling a few hundred ERP records a month typically uses a few US dollars to a few tens of dollars of model time. The larger lines are the ERP API licence, hosting and the person reviewing the drafts.

At a glance

• Who it is for: owners, CFOs and operations managers on Odoo or MYOB Acumatica who have been quoted, or are about to scope, a first AI agent and want the running cost itemised before they sign.
• Problem it solves: every quote covers the build; almost none itemises what the agent costs each month afterwards, so the budget stops at the pilot and the recurring lines arrive as surprises.
• Platforms: Odoo 19 Enterprise (Ask AI, the AI app, JSON-RPC API); MYOB Acumatica 2025.2 and 2026.1 (AI Automation, the REST API); any hosted model from OpenAI, Anthropic or Google, or a self-hosted model, connected through an MCP server or the platform API.
• Typical cost, model usage: for a reconciliation exception agent working 300 exceptions a month, under US$1 to about US$26 a month depending on the model tier, from the worked example below using the Anthropic, OpenAI and Google price lists as at 15 September 2026.
• Typical cost, ERP side: Odoo’s external API is on the Custom plan, A$64.00 then A$80.00 per user per month ex GST on yearly billing (Australian pricelist as at 5 October 2026, odoo.com/pricing), and the agent’s own login is one of those users; MYOB Acumatica includes one API user at 1,500 calls a day and sells unlimited access separately, per MYOB’s API licence page, priced through your partner.
• Typical cost, build: a scoped agent pilot typically starts from a few thousand dollars ex GST for discovery and sandbox build, per our AI agents for ERP in Australia page.
• Typical timeline: 4 to 10 weeks from discovery to a drafts-first go-live, per the same page; the running cost starts the day the agent goes live.
• Last verified: 15 September 2026 against the three provider price pages above, Odoo’s AI API key documentation, MYOB’s LLM provider documentation and the ATO’s page on GST for imported services.

The question arrives about ten minutes into every scoping call. Someone has read that AI models charge by the token, has no idea what a token is, and wants to know whether the agent will quietly run up a bill like a forgotten cloud server. It is the right question. It just has a smaller answer than most people expect, and a larger one hiding behind it.

We run agents in production on our own operations and against client ERPs, so we see the invoices. This is what an AI agent costs to run as it shows up on them: five lines, one worked example, and the controls that keep each line where you put it.

What an AI agent costs to run: the five lines

An AI agent over an ERP has five recurring cost lines. Most write-ups cover the first and stop.

• Model usage. The tokens the agent sends to and receives from the model, billed per million tokens by the provider. Small, and the most talked about.
• ERP access. The licence, plan or API tier that lets the agent read and draft in your system under its own login. Often the largest fixed line, and the one quotes leave out.
• Hosting. The MCP server or orchestrator that holds the agent’s tools, schedules its runs and keeps its logs. A small always-on service, budgeted like any integration server.
• Review time. The minutes a named person spends approving, editing or rejecting what the agent drafts. The real cost, and the one you should never cut.
• Governance overhead. The register entry, the quarterly permission review, the log check. An hour or two a month per agent once the routine exists.

The build and pilot sit in front of all five as a one-off; the AI agents for ERP in Australia page covers what that looks like. Everything below is about what happens after go-live.

How much do tokens cost, and how many does one agent use?

A token is a piece of text, roughly four characters or three quarters of a word in English, per Anthropic’s pricing FAQ. Providers charge separately for input tokens (what you send: instructions, the record, the rules) and output tokens (what comes back: the classification, the draft). Output costs about five times input on every list we checked.

The prices themselves, as at 15 September 2026 and all in US dollars: OpenAI’s API price list runs from US$0.20 per million input and US$1.20 per million output on its small model to US$5 and US$30 on its flagship; Anthropic’s from US$1 and US$5 to US$10 and US$50; Google’s Gemini API lists its current Flash tier at US$0.75 and US$3.75, with a published step up to US$1.50 and US$7.50 from 1 January 2027. That last line matters more than the numbers: prices move, in both directions, and a budget set in September is a guess by March.

Two discounts apply to almost every ERP agent. Prompt caching, where the unchanging part of your prompt (the instructions, the tax code allowlist, the exception rules) is stored and re-read at a tenth of the input price on both the OpenAI and Anthropic lists. And batch processing, a 50% discount on both lists for work that can wait a few hours, which describes most overnight reconciliation and replenishment runs.

How many tokens does one job use? For a reconciliation exception, our runs land around 3,000 tokens of standing instructions, 2,000 to 4,000 tokens of record data (the bank line, the open items, the partner history) and 300 to 700 tokens of output. Call it 6,000 in and 500 out. That is the unit the worked example below is built on.

Which ERP licences and API fees does an agent add?

This is the line that decides the budget, and it is different on each platform.

Odoo 19. The built-in Ask AI and the AI app run under Odoo’s own model connection on Odoo Online, where your own API key is optional; on Odoo.sh and on-premise databases a key from OpenAI or Google is required and the provider bills you, per Odoo’s AI API key documentation. An external agent is a different matter. It reaches Odoo through the JSON-RPC API under its own user, and the external API is a Custom plan feature, A$64.00 per user per month for the first year then A$80.00 at list, ex GST on yearly billing per the Australian pricelist as at 5 October 2026 (odoo.com/pricing). The detail most quotes miss: the agent’s login is an internal user, and Odoo charges for internal users, so budget one seat for the agent itself. Our Odoo pricing in Australia post has the full plan maths.

MYOB Acumatica. The native AI features, including AI Automation prompts, are switched on under Enable/Disable Features and are “subject to licensing” per MYOB’s LLM provider documentation, so the first question to your account manager is whether your edition includes them. On version 2026.1 the MYOB-managed model connection is available immediately with no key to manage; if you prefer your own provider, MYOB supports Azure, Amazon Bedrock, OpenAI and Anthropic connections through its Secure Cloud proxy, and you hold the provider relationship and pay the provider. One line from that page belongs in every budget: providers retire models with little notice, and a retired model breaks the connection until someone updates it. An external agent connects through the REST API, where MYOB’s API licence page includes one named API user at 1,500 calls a day and sells a paid Full Access licence with unlimited calls for business-critical or multiple integrations. An agent that polls for changes every minute uses 1,440 of those calls before it does any work, so most production agents need the paid tier. Its price is quoted through your partner; MYOB does not publish it.

Either platform, through MCP. If the agent talks to the ERP through an MCP server, the server is where the scoped credentials live and where the reads and drafts are logged. It adds the hosting line above and removes nothing from the ERP line. Our MCP for ERP guide covers the setup.

Per-seat assistant or metered API: which costs less for an agent?

Most businesses arrive with a chat subscription already in place and assume the agent can run on it. It usually can’t, and the vendors say so. OpenAI’s pricing page states that API access is billed separately from its ChatGPT plans, and MYOB’s documentation says you cannot create an LLM connection with a ChatGPT account; you need the OpenAI API. Anthropic’s Team plan is US$20 to US$25 per seat per month per claude.com/pricing as at 15 September 2026, and its Enterprise plan is a seat price plus usage at API rates. A seat is for a person at a keyboard. An agent that runs on a schedule at 2 am is a metered API workload.

The pattern we build sits in the third column, and the reason is not price. It is that a scheduled agent needs its own credentials, its own logs and its own spend limit, and a seat gives it none of those.

A worked example: the reconciliation exception agent, costed

A Queensland wholesale distributor runs Odoo 19 with a reconciliation exception agent connected through the API on its own user. Odoo’s own matching rules clear the routine bank lines; the agent works the leftovers, about 300 a month, and drafts a classification and a proposed entry for each. Nothing posts until the accounts receivable officer approves it in the Bank Matching view. Details are changed and combined across engagements, as always.

Per exception: 3,000 tokens of standing instructions (cached), 3,000 tokens of record data, 500 tokens of output. At a mid-tier published rate of US$2 per million input, US$0.20 per million cached input and US$10 per million output, that is US$0.006 for the record data, US$0.0006 for the cached instructions and US$0.005 for the draft, about 1.2 US cents an exception. Three hundred exceptions cost US$3.48 a month. Run the same job on the cheapest tier on the lists above and it is about 54 US cents; on the dearest, about US$25.50. Add a daily summary call and round up: the model line for this agent is under US$30 a month on any published tier, and under US$5 on the tier we would actually pick for it.

The ERP line is bigger and fixed. One Custom-plan user for the agent, A$64.00 a month in year one and A$80.00 from month 13, ex GST. The hosting line is a small cloud instance shared with the business’s other integrations.

Then the line that matters. The AR officer spends about 20 minutes a day on the queue, reading each draft’s reasoning, applying the ones that are right and correcting the ones that aren’t. Twenty minutes a day is roughly seven hours a month. Price those hours at her loaded rate (our cost of an employee calculator will do the maths) and they exceed the model and the licence combined, every month. That is the correct outcome. Those seven hours replace the twenty or so she used to spend investigating the same lines from scratch in the last week before the Business Activity Statement (BAS), and they are the reason the entries that post are right.

“When a prospect asks what the agent costs to run, they expect a token figure. I give them one, and it’s usually less than the coffee budget. Then I tell them the number that matters is the twenty minutes a day their AR officer spends approving drafts, because that’s the line that makes the agent safe. Cut the tokens all you like. Never cut the reviewer.”

Josh Craig, Director, Auboros

How do you set a budget per agent and stop it running away?

Five controls, each cheap, each skipped by someone who later wished they hadn’t.

• One key per agent. Every agent gets its own provider API key and its own ERP user. Shared keys make the invoice unreadable and the permission review impossible; the AI agent register records which key belongs to which agent and who owns it.
• A spend limit at the provider. OpenAI’s pricing page describes a monthly budget in billing settings after which requests stop, plus an email threshold; Anthropic’s Enterprise plan lets administrators set user and organisation spend limits, and its console tracks usage. Set the limit at twice the pilot’s measured monthly cost, and treat hitting it as an incident to investigate, not a reason to raise it.
• A cap in the orchestrator. Records per run, runs per day, tokens per record. A loop that re-reads the same 4,000 open items every minute because a filter broke is the failure mode that produces a surprising bill, and it is stopped by a counter, not a policy.
• A cost column in the register. Budget, actual, and the date it was last checked, per agent, per month. The governance controls that make an agent safe also make it cheap to audit.
• A model review date. Providers retire models, as MYOB’s documentation warns and as Anthropic’s price list shows with its retired rows. Put a quarterly date in the register to check the model is still offered and still the right price, and update the connection before it breaks.

Two Australian details for the finance team. The provider invoices are in US dollars, so the AUD figure moves with the exchange rate and the accounts payable process needs a foreign-currency supplier. And GST: if your business is registered for GST and you give the overseas provider your ABN and state that you are registered, GST should not be charged on the purchase, per the ATO’s guidance on imported services and digital products; the reverse charge applies only where you could not have claimed a full GST credit, and the supplier is not required to issue a tax invoice. Talk to your accountant about how that lands in your BAS.

Where to cut cost without cutting the approval gate

Every optimisation we make sits on the model and hosting lines, and none touches the reviewer.

Cache the standing instructions, because the rules do not change between exceptions and the cache read is a tenth of the price. Batch the overnight runs, because a reconciliation that posts at 7 am does not care whether the model answered at 2 am or 4 am, and both major lists give the batch discount at half price. Use a small model for the classification step and a larger one only for the draft that a person will read; the classification is a sorting job, the draft is a writing job, and they do not need the same brain. Trim the record data to the fields the rule actually uses; the partner’s full history is rarely needed when the last five payments will do.

What we do not do is widen the approval threshold to save review minutes. An agent that posts small write-offs on its own because it was right last quarter has traded a few hours of an AR officer’s time for an unreviewed ledger, and no token saving pays for that. The cost of an agent is a token bill you can round to zero, a licence you can read off a price list, and a person’s attention you should budget properly and protect. Get the third one right and the first two take care of themselves.

Goes well with

• AI agents for ERP in Australia: what the build and pilot cost before the running cost starts, and which agent to scope first.
• MCP for ERP: how the scoped keys, the hosting line and the logs fit together when the agent connects through an MCP server.
• The AI agent register: the one-page record where the cost column, the key owner and the model review date live.

Budgeting your first agent and want the recurring lines itemised?

Auboros designs and runs drafts-first AI agents over Odoo and MYOB Acumatica from Brisbane, and every proposal we write itemises the model, licence, hosting and review lines for the first year, not just the build. If you want to know what an agent would cost to run on your volumes, book a free consultation. Bring a month of the work you want automated and we’ll cost it line by line, including the part that should stay human.